CN113892148A - Interpretable AI (xAI) platform for computational pathology - Google Patents

Abstract

Pathologists use digital pathology for diagnosis by using full-section images (WSI). Interpretable AI (xai) is a new approach to AI that can reveal the root cause of its results. It can thus be seen that xAI can improve the safety, reliability, and interpretability of machine learning for critical tasks such as pathological diagnosis. HistoMapr provides the pathologist with intelligent xAI guidance to improve the efficiency and accuracy of pathological diagnosis. HistoMapr may preview WSI for an entire pathological case, identify critical diagnostic regions of interest (ROIs), determine one or more conditions associated with each ROI, temporarily mark each ROI with the identified conditions, and may shunt them. The ROI is presented to the pathologist in an interactive, interpretable manner for rapid interpretation. The pathologist can control and can go through "why "interface access xAI analysis. HistoMapr can track the pathologist's decisions and compile pathology reports using suggested, standardized terms.

Description

Interpretable AI (xAI) platform for computational pathology

Cross Reference to Related Applications

Priority and benefit of U.S. provisional patent application No. 62/819,035 entitled "An extensible AI (xAI) Platform for computerized Pathology," filed on 3, 15, 2019, which is incorporated herein by reference in its entirety.

Technical Field

The present disclosure relates generally to Artificial Intelligence (AI) -based systems that perform image analysis and processing, and more particularly to a system that can assist in classifying one or more regions of an image based on respective conditions indicated by the regions, where the conditions are determined using an interpretable artificial intelligence (xAI) -based platform.

Background

With the recent development of digital pathology full-slice image (WSI) platforms, more and more pathologists have transitioned to viewing digital images of patient slices on computer monitors. After a long time course, the FDA recently approved the first WSI system as a class II medical device, and there are several additional WSI platform vendors (e.g., lycra (Leica), Philips (Philips), Roche (Roche), etc.). In response to recent technological advances beyond these new market forces and pathologies, new areas of applying machine learning to the computational pathology of WSI are emerging. This has also led to innovative new concepts, such as computer aided diagnosis (pCAD) for pathologists, which suggests a method to integrate machine learning into the pathology workflow.

Deep learning in the form of convolutional neural networks was prevalent in early computational pathology work. Although powerful in isolated, lower-level applications such as mitotic counting or cancer detection, deep learning has not resulted in validated, comprehensive, advanced systems. There is also fear and doubt of pathologists applying AI to pathology and there is no consensus on how a pathologist should supervise or work with a computational pathology system.

Disclosure of Invention

To address the above challenges, various techniques described herein use xAI to analyze and infer pathology. As an example, xAI was used to assist a pathologist in deriving inferences for a breast biopsy. The key new concept is interpretable AI (xai), which refers to an AI whose results can be proven by data. This is intended to address certain issues related to AI-based analysis, such as bias, transparency, safety, and causality. Bias refers to potentially defective AI caused by biased training data that may initially appear to be functional, but may later fail catastrophically. In various embodiments, our xAI platform (also known as HistoMapr) may prevent such events by providing its users with real-time feedback on the process of deriving inferences of xAI. Using this feedback, a user (e.g., a pathologist) may determine xAI whether to infer a training data set based on the deviation. The transparency is relevant; xAI can prove to the user that the results are correct, which provides them with all the information needed to make a good decision based on xAI recommendations.

In various embodiments, such information may include:

1) why-a list of quantitative features, each feature having a positive/negative score that affects both the diagnostic label and its confidence score;

2) why not-the same as above, but also an ordered list of alternative provisional diagnoses with its own positive/negative record of quantitative characteristics. The features associated with the alternative provisional diagnoses are the same as described above, but they typically have different positive/negative values for the various provisional diagnoses;

3) i know where you succeed-in the context of the ROI image, in the context of the domain knowledge of the pathologist user, features in 1 and 2; if consistent, the user is assured that HistoMapr was successful

4) When you failed-the same features as #3, but if they did not agree, the pathologist knows xAI that it is likely that it has failed

5) When to trust-based on full situation awareness of why/why failure, success/failure for all ROIs

6) When not trusted-an execution decision based on the aforementioned factors; the pathologist may be offered the option of distrust xAI and proceed to review for diagnosis with manual WSI.

Patient safety is paramount in pathology and is a result of complex interactions between pathologists, other physicians, and laboratory personnel and computer systems including computer pathology applications. Patient safety can be improved by xAI (not only by reducing undetected deviations or by providing transparency to the pathologist), which also relates to the ability of the pathologist to monitor xAI the individual patient samples in real time. In other words, various embodiments of the xAI system described herein provide an explanation of a pathology in a particular individual as the pathologist reviews the pathology. The interpretation is not limited to retrospective or cumulative data audits for multiple patients.

Finally, causal relationships refer to scientific understanding of the pathological mechanisms underlying the xAI system. Many in-depth learning efforts merely pursue engineering statistics (e.g., area under the curve, AUC) without gaining new insight into the pathology itself. However, by disclosing quantitative features, the HistoMapr embodiment reveals the pathological mechanisms underlying the decision making by the xAI system. It follows that embodiments of HistoMapr may allow researchers to understand new disease mechanisms that may lead to meaningful diagnostic or therapeutic advances.

Accordingly, in one aspect, a method for performing interpretable pathology analysis of a medical image is provided. The method comprises the following steps: for a region of interest (ROI) in a full-slice image (WSI) of tissue, features of several different feature types are identified, wherein at least one feature type is at least partially indicative of a pathological condition of tissue within the ROI. The method further comprises the following steps: classifying the ROI into one of a plurality of tissue condition classes; and (ii) assigning a label to the ROI that indicates a condition of tissue associated with the category, wherein the label will then identify the condition of tissue in the ROI. For classification, a classifier trained to classify using features of the identified feature type is used to classify the tissue image into one of several tissue condition classes.

In addition, the method comprises: explanatory information about the designation of the tag is stored. The explanatory information includes information about the identified features. Further, the method comprises: displaying: (i) at least a portion of the WSI in which the boundary of the ROI is highlighted; (ii) a label assigned to the ROI; and (iii) a User Interface (UI) having: (a) a first UI element for providing a user with access to the stored explanatory information; and (b) one or more additional UI elements that enable the user to provide feedback on the specified tags.

In some embodiments, the tissue comprises breast tissue and the categories of tissue conditions include two or more of: invasive cancer, Ductal Carcinoma In Situ (DCIS), high risk benign, low risk benign, ductal epithelial atypical hyperplasia (ADH), Flattened Epithelial Atypical (FEA), Columnar Cell Change (CCC), and normal duct. In some other embodiments, the tissue comprises lung tissue and the categories of tissue conditions comprise: idiopathic Pulmonary Fibrosis (IPF) and normal. In some embodiments, the tissue comprises brain tissue and the categories of tissue conditions comprise: classical cell tumors and protoneurocyte tumors.

The different feature types may include cytological features and/or Architectural Features (AF). Cytological features may have subtypes such as: nuclear size, nuclear shape, nuclear morphology, or nuclear texture. Architectural features may have subtypes such as: an architectural feature (AF-C) based on the colors of a set of superpixels in the ROI; (ii) architectural features (AF-N) based on the cytological phenotype of nuclei in the ROI; or (iii) a combined architectural feature (AF-CN) based on both the color of a set of superpixels in the ROI and the cytological phenotype of nuclei in the ROI. Architectural features may also have subtypes, such as: nuclear arrangement, stromal cellularity, epithelial pattern in ducts, epithelial pattern in glands, cell cobblestone, stromal density, or hyperplastic.

In some embodiments, the information about the features includes one or more of the following: a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label; a count of features of a particular feature type detected in the ROI; a measured density of features of a particular feature type in the ROI; or the strength of a particular feature type in indicating a condition of the tissue.

The explanatory information may include a confidence score calculated by the classifier in specifying the label. The confidence score may be based on one or more of the following: a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label; for a first feature type: (i) a strength of the first feature type in indicating a condition of the tissue; or (ii) a count of features of a first feature type detected in the ROI; or another total number of feature types that are detected in the ROI but that correspond to a different tissue condition than the condition associated with the label.

In some embodiments, the method further comprises: in response to a user interacting with the first UI element, an explanatory description is generated using the standard pathology vocabulary and stored explanatory information, and the explanatory description is displayed in an overlay window, side panel, or page. The method may further comprise: highlighting a feature of the particular feature type in the ROI using a color assigned to the particular feature type, the feature being at least partially indicative of the tissue condition indicated by the label; and displaying the highlighted ROI in the overlay window, side panel, or page.

In some embodiments, a method comprises: the identifying, designating and storing steps are repeated for a plurality of different ROIs. In these embodiments, the method further comprises, prior to the displaying step: (i) calculating a respective risk metric for each ROI, wherein the risk metric for the ROI may be based on: (a) a designation label of the ROI; and/or (b) a confidence score for the ROI; and (ii) ordering the ROIs according to their respective risk metrics. In these embodiments, the displaying step comprises: displaying in one panel: (i) highlighting at least a portion of the WSI that has the boundary of the ROI with the highest risk metric; (ii) a label assigned to the ROI; and (iii) a User Interface (UI) that provides the user with access to the stored interpretation of the specified label for the ROI. The displaying step further comprises: thumbnails of the ROI sequence are displayed in another panel.

In some embodiments, a method comprises: obtaining a full slice image (WSI); and identifying the ROI in the WSI. The identification of the ROI may include: (i) marking at least two types of superpixels in the WSI, wherein one type corresponds to hematoxylin stained tissue, and the other type corresponds to eosin stained tissue; and (ii) marking segments of the first type of superpixel to define the occlusion region as a ROI. The method can comprise the following steps: several different ROIs in the WSI are identified.

In some embodiments, a method comprises: the training data set for the classifier is updated. Updating the training data set may include: receiving feedback from the user via one or more additional UI elements for the label designated to the ROI, wherein the feedback indicates a correctness of the designated label as perceived by the user; and storing the portion of the WSI associated with the ROI and the specified label in a training dataset. The classifier may be a decision tree, a random forest, a support vector machine, an artificial neural network, or a logistic regression-based classifier.

In another aspect, a method for distributing cases among a group of pathologists is provided. The method comprises the following steps: for each of several cases, processing a corresponding full-slice image (WSI) of the tissue, wherein the processing of the WSI comprises: one or more regions of interest (ROIs) in the WSI are identified, wherein each ROI is assigned a respective diagnostic label indicative of a condition of tissue in the ROI. The diagnostic label may be generated by the classifier or may have been provided by the pathologist.

The method further comprises the following steps: for each ROI, calculating a respective confidence score for the respective designation; for WSI calculation: (i) a severity score based on the respective diagnostic label assigned to one or more ROIs in the WSI; and (ii) a confidence level based on the respective confidence scores of the one or more ROIs. In addition, the method comprises: the severity score, confidence level, and corresponding confidence score are stored as explanatory information. The confidence score for the ROI may be calculated for tags suggested by the platform or for tags provided by the pathologist.

The method further comprises the following steps: it is determined whether the severity score is equal to or above a specified threshold severity score and, if so, the WSI is sent to an emergency pathologist in a group of pathologists who may review the WSI immediately (or with high priority). The method comprises the following steps: if the severity score is below a specified threshold severity score, determining whether the confidence level is equal to or below a specified threshold confidence level, and if so, sending the WSI to a second specialist in the pathologist group; otherwise, the case is sent to the general pathologist in the group of pathologists. Thus, the second specialist can review relatively difficult cases with high priority. Sending the case to the general pathologist may include: a general pathologist is selected from a library of general pathologists within a group of pathologists such that a balanced workload of the library is maintained when sending cases to the selected general pathologist.

In some embodiments, a method comprises: assigning a corresponding diagnostic label to at least one ROI in at least one WSI. Assigning a corresponding diagnostic label to a particular ROI includes: using a classifier (trained to classify an image into one of a plurality of tissue condition classes using features of several different feature types identified in the image) to: classifying the ROI into a category within a plurality of tissue condition categories; and assigning a label to the ROI that indicates a condition of tissue associated with the category. In some other embodiments, for at least one ROI in at least one WSI, the corresponding diagnostic label is provided by a previous reviewer and a pathologist group represents a subsequent reviewer group.

In some embodiments, the method further comprises: in response to a user requesting an interpretation of a particular WSI's transmission via a UI element: an explanatory description is generated using the standard pathology vocabulary and the stored explanatory information, and the explanatory description is displayed. The method may further comprise: selecting an ROI from a particular WSI for which the specified label indicates a severe condition or for which the confidence score is equal to or below the specified threshold confidence score. Additionally, the method may include: highlighting a feature of a particular feature type in the ROI using a color assigned to the feature type, the feature type being at least partially indicative of a tissue condition indicated by a label assigned to the ROI; and displaying the highlighted ROI along with the explanatory description.

For a first ROI in a first WSI, the explanatory information may include one or more of the following information: a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI; a count of features of a particular feature type detected in the first ROI; a measured density of features of a particular feature type in the first ROI; or the strength of a particular feature type in indicating a corresponding tissue condition.

The confidence score for the first ROI in the first WSI may be based on one or more of: a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI; for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of a first feature type detected in the ROI; or another total number of feature types that are detected in the first ROI but that correspond to a different tissue condition than the condition associated with the label assigned to the first ROI.

In another aspect, a method for truth labeling an image used to train a classifier is provided. The method comprises the following steps: obtaining a full-section image (WSI) of the tissue; and identifying one or more regions of interest (ROIs) in the WSI. The identification of the ROI includes: (i) marking at least two types of superpixels in the WSI, wherein one type corresponds to hematoxylin stained tissue, and the other type corresponds to eosin stained tissue; and (ii) marking segments of the first type of superpixel to define the occlusion region as a ROI. The method further comprises the following steps: displaying one or more ROIs in sequence; and for each ROI: one or more UI elements are displayed, wherein a first UI element may provide or confirm a respective truth label to be assigned to the ROI. In addition, the method comprises: responsive to a user interacting with the first UI element, a corresponding truth label is specified to the ROI and the ROI is stored in a training corpus.

In some implementations, the first UI element indicates that the user agrees to the provided suggestion, and the method includes: for each ROI of at least a subset of the one or more ROIs: features of several feature types are identified, wherein at least one feature type is at least partially indicative of a pathological condition of tissue within the ROI. The method further comprises the following steps: using a classifier (trained to classify an image into one of a plurality of tissue condition classes using features of several different feature types) to: (i) classifying the ROI into a category within a plurality of categories; (ii) assigning a suggestion tag to the ROI that indicates a condition of tissue associated with the category; and (iii) storing explanatory information regarding the designation of the suggested tag. The explanatory information includes information about the identified features. The method further comprises the following steps: the suggested tags are displayed as the provided suggestions.

The method may further comprise: in response to a user requesting interpretation of a suggested tag for a particular ROI via a UI element: an explanatory description is generated using the standard pathology vocabulary and the stored explanatory information, and the explanatory description is displayed. In some embodiments, a method comprises: highlighting a feature of a particular feature type in a particular ROI using a color assigned to the feature type, the feature type being at least partially indicative of a tissue condition indicated by a label assigned to the ROI; and displaying the highlighted ROI along with the explanatory description.

For a particular ROI, the explanatory information may include one or more of the following information: a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI; a count of features of a particular feature type detected in a particular ROI; a measured density of features of a particular feature type in a particular ROI; or the strength of a particular feature type in indicating a corresponding tissue condition.

The confidence score for a particular ROI may be based on one or more of the following: a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI; for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of a first feature type detected in a particular ROI; or another total number of feature types that are detected in the particular ROI but that correspond to a different tissue condition than the condition associated with the suggestion tag assigned to the particular ROI.

In another aspect, a system is provided for: (i) performing an interpretable pathology analysis of the medical image; (ii) distributing cases among a group of pathologists; or (iii) true value labeling of the images used to train the classifier. The system includes a first processor and a first memory in electrical communication with the first processor, wherein the first memory includes instructions executable by a processing unit, which may include the first processor or a second processor, and in electrical communication with a storage module, which may include the first memory or the second memory. The instructions program the processing unit to perform one or more operations and/or one or more steps thereof according to the method described above.

The various embodiments described herein have several technical effects. For example, the accuracy of image analysis of a full-slice image of tissue can be greatly improved because the process of identifying a region of interest in the full-slice image scans the entire image without omitting any portion thereof, and labels any region that may be associated with a diagnostic condition as a region of interest. Thereby, the probability of missing a region in the full-slice image having a condition that should be diagnosed is minimized, which may minimize errors of the missed diagnosis, thereby improving the quality of care by minimizing the risk of diseases, severe health conditions, or complications that may be caused by the missed diagnosis.

Moreover, because the various embodiments described herein suggest not only a provisional diagnosis, but also an explanation of how to make a particular diagnosis determination, the embodiments may improve the accuracy of the diagnosis and may also improve consistency between auditing pathologists. In addition, to suggest a provisional diagnosis, various embodiments employ an artificial intelligence based system that can be continuously trained, especially when having images that are inconsistent between the provisional diagnosis and the final diagnosis performed by the reviewer. In some embodiments, the system itself facilitates generating a truth-labeled data set that may be used for initial and/or ongoing training. This may further improve the accuracy of the diagnosis.

Some embodiments may improve the accuracy of the diagnosis in yet another way. In particular, these embodiments may assign cases to pathologists based on their expertise, for example, in their reviewers. This assignment may be performed as part of the final determination of the initial diagnosis or after the final diagnosis for quality checking and assurance. Assigning cases to reviewers based on the reviewer's expertise may also improve diagnostic accuracy, as specialists or sub-specialists are not burdened with having to review simple cases and may focus on difficult to diagnose cases. Also, general practitioners (generalist) or less experienced reviewers are not burdened with reviewing difficult to diagnose cases, as their review may be erroneous or they may not be able to review such images at all. Moreover, cases that require immediate attention because the underlying condition is severe can be given such attention by directing them to the auditor responsible for auditing the cases, which can also improve the quality of care.

Drawings

The patent or application file contains at least one drawing executed with shading. Copies of this patent or patent application publication with color drawing(s) will be provided by the office upon request and payment of the necessary fee.

The present disclosure will become more apparent in view of the attached drawings and accompanying detailed description. The embodiments depicted therein are provided by way of example and not limitation, wherein like reference numerals/signs generally refer to the same or similar elements. However, in different drawings, the same or similar elements may be referred to using different reference numerals/numbers. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating aspects of the invention. In the drawings:

1A-1C schematically depict the use of an xAI-based platform in pathology, according to various embodiments;

FIG. 2A illustrates a conventional process of analyzing a full-slice image;

fig. 2B illustrates a typical pathology analysis that may be performed using an embodiment of the xAI platform;

FIG. 3 illustrates the natural hierarchy found in the spatial organization of an organization, where such hierarchy is used for analysis performed by various embodiments of the xAI platform;

FIG. 4 illustrates a pipeline of analysis performed by the xAI platform according to some embodiments;

fig. 5 illustrates a pathology analysis that may be performed using an embodiment of the xAI platform;

FIG. 6 illustrates an example of displaying information that justifies the analysis performed by the xAI platform, according to one embodiment;

FIG. 7 illustrates examples of image features relating to different tissue conditions, where such features are used by embodiments of the xAI platform;

FIG. 8 illustrates an example Graphical User Interface (GUI) that may be used for truth labeling to generate a training corpus for training xAI embodiments of the platform;

FIG. 9 illustrates another example of displaying information that justifies the analysis performed by the xAI platform, according to one embodiment;

FIG. 10 illustrates yet another example of displaying information that justifies the analysis performed by the xAI platform, wherein the displayed information includes regions of interest similar to the regions of interest being analyzed, according to one embodiment;

FIG. 11 is a flow diagram of the offloading performed by the xAI platform, according to some embodiments;

FIG. 12 compares pathology analysis performed according to the conventional technique with analysis performed via various embodiments of the xAI platform;

FIG. 13 shows a table of categories of organization conditions that may be used by embodiments of the xAI platform;

FIG. 14 illustrates an example GUI that allows a user to diagnose and correspondingly label tissue images, wherein no suggestions are provided regarding the condition of the tissue;

FIG. 15 illustrates another example GUI that allows a user to diagnose and label tissue images accordingly, wherein suggestions regarding tissue condition are provided by the xAI platform, according to one embodiment;

fig. 16 illustrates examples of various architectural features of breast tissue and different breast tissue samples having different architectural features, which some embodiments of the xAI platform may use for analyzing breast tissue samples;

fig. 17 schematically shows case diversion according to the conventional technique;

fig. 18 schematically illustrates case diversion using the xAI platform, according to one embodiment;

fig. 19 shows an example of a GUI for case diversion that an embodiment of the xAI platform may provide;

fig. 20 illustrates an environment in which an embodiment of the xAI platform may perform case diversion;

21A-21F illustrate a process of identifying features of a tissue image, where the features are used to determine a condition associated with the tissue, and where the determination may be explained using the features using an embodiment of the xAI platform;

22A 1-22A 12 illustrate examples of different lung tissue samples having different architectural features;

FIG. 22B illustrates different architectural features of lung tissue;

FIG. 22C illustrates an image of a lung tissue sample with certain types of features that are too dense, where the tissue image is marked to highlight those types of features, according to some embodiments;

FIG. 22D illustrates an image of another lung tissue sample with some other types of features that are too dense, where the tissue image is marked to highlight those other types of features, according to some embodiments;

fig. 23A to 23D show examples of images of brain tissue samples;

FIG. 24A illustrates different architectural features of brain tissue;

FIG. 24B shows an image of the brain tissue sample shown in FIG. 23A and having certain types of features that are too dense, wherein the tissue image is marked to highlight those types of features, according to some embodiments;

fig. 24C shows an image of another brain tissue sample shown in fig. 23C and having certain other types of features that are too dense, where the tissue image is marked to highlight those other types of features, according to some embodiments; and

fig. 25A-25F illustrate a process of identifying a region of interest from a full-slice image of tissue for further analysis thereof, according to some embodiments.

Detailed Description

Pathology is considered the gold standard for medical diagnosis. Thus, pathologists have been conservative in making large practical changes. Computational pathology is of considerable interest, but is also a concern, as any error made by a computer can cause harm to humans and even a serious risk of loss of life. It is likely that xAI may facilitate adoption of computational pathology, as it not only presents results efficiently, but also explains how the results are obtained. A person, typically a pathologist, can then easily determine whether the computer-generated results should be accepted or whether further review is required. The interpretable AI can also assist in driving a scientific understanding of the pathology underlying cancer and other difficult to understand diseases. It is important to note that xAI is used by us to support pathologists in making efficient and more accurate "calls"

With reference to fig. 1A-1C, the disclosure of information leading to decisions made by embodiments of the xAI platform may make such a platform a key feature of computational pathology. Generally, in various embodiments, HistoMapr xAI may produce content-based image applications that are based on analysis of ROI of interest. HistoMapr may analyze WSI to find one or more ROIs and then provide diagnostic labels to the respective ROIs. HistoMapr may also generate descriptive information about deviations, transparency, safety, causality, etc. HistoMapr may be used by a pathologist to aid diagnosis, where descriptive information may provide real-time interpretation of HistoMapr analysis. In some implementations, content-based image searching is also implemented, which facilitates several related activities, such as decision support, education, or clinical trial screening (lower center). Finally, quality assurance activities can be facilitated and include continuous monitoring by HistoMapr and pathologists, as well as improved reporting through the promotion of standardized terminology.

Referring to fig. 2A, conventional artificial pathology diagnosis, whether using microscope slides or using artificial WSI, is generally inefficient and error prone. Fig. 2A shows a simplified, normal microscope or artificial WSI 202 and a typical observation path 204 (dashed and solid black arrows). The pathologist must observe the tissue in a systematic way to ensure that all tissue is observed. This may lead to efficiency problems where the diagnostic region (circle 206) is found unpredictably and without warning. The diagnostic region may even be missed if small or imperceptible. After making a decision to identify a diagnostic region (or ROI) of interest, a continuous search for potential diagnostic regions is also slower than a confident audit. Finally, there is no prior knowledge as to whether individual pathological cases will be difficult or require acceleration.

It can be seen that when diagnosing difficult lesions such as ductal epithelial dysplasia in breast biopsies, one group reported low diagnostic concordance (52% concordance) between pathologists. The report also notes that different kinds of pathologists (i.e., mastopathologists versus general pathologists) experience different levels of performance. Regarding efficiency, manual WSI observation does not provide sufficient efficiency to justify its implementation for preliminary diagnosis compared to conventional slide microscopy, given the cost of WSI and variations in workflow, including time delays in slice scanning. This is based on unpublished data, but was also confirmed by the slow adoption of digital pathology for preliminary diagnosis in recent years, despite open regulation in 2017.

Fig. 2B depicts an example display 250 of an embodiment of HistoMapr, referred to as HistoMapr-break, configured to analyze Breast tissue WSI. The central panel 252 shows a WSI viewing region with a computer-identified ROI boundary 254 green. Other colors or techniques may be used to generate the boundary. The panel 256 on the right shows the WSI 258 that has been analyzed and is available for viewing. The left panel 260 shows patient information and the provisional diagnosis generated by HistoMapr-break. Bottom panel 262 shows key elements of HistoMapr, including "why? ", the UI elements (e.g., buttons) of xAI 264, the HistoMapr diagnostic tab 266" ADH "(ductal epithelial dysplasia), pathologist UI elements (e.g., buttons) 268a-c indicating consent, disagreement, or uncertainty about the HistoMapr tab, and the completion UI element (e.g., button) 270. Below the indicators and buttons, there is an ROI sub-panel 272 showing ROIs 274 that are shunted from most atypical to least atypical. The ROI 274 is generally shown at a smaller size than the ROI shown in the central panel 252. The ROI 274 may also be shown as a thumbnail image.

According to our simulation of the pCAD model, computational pathology, such as that facilitated by HistoMapr-break, can be 56% more efficient than traditional microscopy for Breast core biopsy. Efficiency gains appear to come from several factors, including early discovery of diagnostic regions of interest (ROIs); reduced uncertainty due to shunting effects; and the ability to review fewer diagnostic ROIs in an accelerated manner after making primary diagnostic decisions.

Defining:

as used in the following discussion, the following terms may generally be interpreted according to their ordinary and general meaning in the context of the field of computational pathology. The following description of various terms is not necessarily limiting, but may convey its ordinary and general meaning.

Pathological cases-this is a patient sample (such as a breast core biopsy) or collection of samples (such as a mastectomy with a single lymph node biopsy) submitted by a physician to a laboratory. These samples or cases are described and processed into microscope slides, which can then be imaged as a full-slice image (WSI). The pathologist observes the microscope slide and/or WSI to make a diagnostic decision, which is then issued as a result report.

Digital pathology-this refers to the use of digital imaging in pathology. Initial efforts focused on remote observation of microscopy using manual and robotic telemicroscopes, or basic image analysis including biomarker readings. Full-section images (WSI) are a more recent field of digital pathology, both for early applications (e.g. telepathology or image analysis) and for more recent preliminary diagnostic applications (i.e. replacement of microscopes). Computational pathology is a new form of digital pathology.

Computational pathology-this is the application of computational techniques including machine learning or AI to the pathology data. This includes image data such as WSI and non-image data such as patient demographics, clinical information, or observations of pathologists.

Artificial Intelligence (AI) -the ability of a computerized system to perform tasks typically associated with a natural biological organism. The term is often applied to projects that develop systems with human mental process characteristics, such as reasoning, finding meaning, summarizing, or learning from experience.

Machine learning-the computer discipline in artificial intelligence research, which mainly involves the implementation of computer software that can be learned autonomously. Expert systems and data mining programs are the most common applications that use machine learning to improve algorithms.

Deep Learning (DL) -part of a broader family of machine learning methods based on learning data representations as opposed to task specific algorithms. Deep learning algorithms mimic the human brain's work in processing data and creating patterns for decision making by utilizing hierarchical levels of artificial neural networks. Deep learning enhanced programs are commonly referred to as "black boxes" in machine learning, where even their designers cannot explain why AIs reach a particular decision.

Computer-aided diagnosis for pathologists (pCAD) -a conceptual framework that lays out an advancement path for applying computational pathology to real-world practice. The intent is to enhance the pathologist with intelligent computer guidance that allows the pathologist to delegate everything that is possible so that they can focus on the critical decisions that only the pathologist can make. This requires a high integration of pathology case information, highly detailed clinical situation planning, and computational pathology pipelines. HistoMapr represents an early iteration of this type of intelligent guidance.

point-by-Point Mutual Information (PMI) -a two-dimensional map defined by the total number of cellular phenotypes used to record the relative co-occurrence and anti-correlation of spatially distributed cellular phenotypes in tissue samples. PMI maps with strong diagonal elements and weak off-diagonal elements describe tumor samples that are locally homogeneous and globally heterogeneous with respect to the spatial distribution of cell phenotype. PMI maps with strong off-diagonal elements describe tumors with many local interactions between different cell phenotypes, thus suggesting that tumor samples exhibit strong local heterogeneity.

Interpretable artificial intelligence (xAI), also known as transparent AI or interpretable machine learning, is artificial intelligence where actions can be trusted and easily understood by humans. xAI the algorithm is programmed to describe its purpose, rationale and decision-making process in a manner that is understandable to the ordinary person. xAI are generally discussed in terms of deep learning and play an important role in the FAT ML model (fairness, responsibility, and transparency in machine learning). Although the term xAI is used generally to explain the black box DL tool, it is not limited to the FL tool and applies generally to machine learning, including AI algorithms, and can generate human understandable statements describing inferences and intermediate and final conclusions of the machine learning/AI process.

HistoMapr and xAI

To address the unmet need for efficiency and accuracy in pathology diagnostics, and to speed up the adoption of computational pathologies in pathology practice,we have developed a so-called HistoMapr^TMxAI platform. HistopMapr-Breast, described above with reference to FIG. 2B, is an example of a HistopMapr platform. In various embodiments, HistoMapr reflects the natural hierarchy found in the spatial organization of human tissue, as shown in the pyramid diagram shown in fig. 3. At the lowest level, simple image analysis is performed, such as cell nucleus size/shape determination, mitotic counting, color optimization, and the like. This analysis can be integrated with a point-by-Point Mutual Information (PMI) map of the second layer to identify and classify potential features in the WSI, for example, tissue structures such as ducts, inflammatory infiltrates, or tumors in breast tissue and their locations within a region of interest (ROI).

At the third level, diagnostic markers for regions of interest (ROIs) are identified, wherein the markers are based on spatial relationships of tissue structures and/or cell populations, and the ROIs are also marked with such diagnostic information at layer 3/level. The labeled ROI may then be shunted based on the diagnostic meaning at layer 4/level. Thus, the xAI HistoMapr guide appears; WSI is represented as a guided review of shunted ROIs in the context of a pathologist's diagnostic task. Additional details regarding the various steps in the process and the manner in which the analysis is performed at the various levels are provided in appendix A below.

Various embodiments of HistoMapr enable precise medical methods to be incorporated into routine diagnostic and prognostic activities through relationships between different cell populations including tissue microenvironments (e.g., tumors, stromal cells, and immune cells). As an example, a pathologist may have difficulty consistently diagnosing breast core biopsies. Thus, we created an embodiment of HistoMapr-Breast (discussed above with reference to fig. 2B), which analyzes the WSI of the mammary gland. HistoMapr-break is able to analyze whole Breast core biopsies, which may include one or more WSIs, using basic image analysis and PMI maps to locate a region of interest (ROI) containing a diagnostic Breast tissue structure, such as a duct, vessel, or stromal infiltration.

In particular, fig. 4 shows a computational pathology pipeline for an embodiment HistoMapr-Breast featuring an end-to-end high risk benign Breast lesion detector for WSI. In step 404, from the WSIs received in step 402, a WSI is selected and analyzed. In particular, WSI stain color normalization is performed in sub-step (a), followed by catheter ROI segmentation in sub-step (B). Cytological and architectural feature extraction (phenotypic analysis) is performed in sub-steps (C) and (D), respectively, which results in a catheter ROI classification in step 406, where the respective regions may be labeled or identified as high risk or low risk.

In some embodiments, HistoMapr-break analyzes the ROI to find pre-trained features and quantify them (if present). This analysis can be used to mark the ROI with diagnostic terms such as "ductal epithelial dysplasia", "invasive cancer", and the like. Such labeling is based on an analysis of the pattern and intensity of architectural features, where intensity may indicate how strongly a particular feature differs. HistoMapr may also use a confidence score that combines features and feature quantities to indicate its confidence in the tag. The labeled ROIs can then be shunted based on both the diagnostic labels and the confidence scores of those labels.

For example, the ROI may be classified from benign to malignant, or from benign to atypical if cancer is not present. Within the diagnostic category, the ROI may be shunted based on the confidence score. These steps occur before the pathologist begins to observe the case, possibly during the night or during weekend off-hours. Using a pathologist-centric interactive interface, HistoMapr can display the ROIs in shunt order so that the pathologist sees the most malignant or atypical regions first (if any) as shown in fig. 5 (same as fig. 2B). In particular, in some embodiments, the ROIs are ranked using a risk metric in the following manner: the high- risk ROIs 502a, 502b are shown first to the pathologist, followed by the lower- risk ROIs 504a, 504b in the bottom sub-panel 272 of HistoMapr-break. The ROI associated with the highest risk is shown first in the central panel 252. Critically, at any time, the pathologist may have full control and may assume manual control of the WSI viewer system. The risk metric for a particular ROI may be based on the severity of the label assigned to the ROI, the confidence score corresponding to the marker, the difficulty or uncertainty of deriving the ROI, or a combination of two or more of these factors.

Referring to fig. 6, the xAI aspect of various embodiments appears as "why? "button that provides one or more panels of supplemental information on how the AI reaches a particular decision. The pathologist can click on "why? The "or" explain "button 604 accesses the interface panel 602 or page (not shown) to interpret xAI the output. The display panel 606 of the HistoMapr-Breast xAI system shows a region of interest 608 with green borders that is classified as "ADH" 610. The click name is "why? "button 604 opens an explanation panel 602 entitled" key findings "in which the reason from the ROI in question is shown to the pathologist as the reason the xAI system decided to classify the region as ADH. Typically, these reasons are based on information about features identified in the ROI. The information includes the type or kind of features identified, their strength in determining the selected marker, the count of different types of features, the type and/or count of features that imply different markers, and the like. The reasons may be displayed using feature information and standard vocabulary commonly understood by pathologists.

In the examples discussed herein, the reasons include: (i) strong evidence of rigid architecture; (ii) determining that the detected tissue is non-hyperplastic; and (iii) have a very high matrix density. In some cases, the strength of the stiffness may be quantified, for example, on a numerical scale or as a percentage relative to a norm. Likewise, matrix density can also be quantified. Additionally, a confidence score may also be presented, which is determined using the t-value or z-value and indicates xAI the confidence in labeling the ROI.

Thus, the pathologist may have a complete situational awareness, thereby being able to make very optimal diagnostic decisions. The HistoMapr embodiment also facilitates the pathologist's work by managing diagnostic information and tracking the pathologist's consent or non-consent to the provided diagnostic label. The pathologist may also indicate uncertainty, and some embodiments of HistoMapr collect this information for possible additional staining disease review or consultation. When the pathologist is ready to complete a case, he or she can press the "complete" button 612 and, in response, HistoMapr can automatically construct a results report using the pathologist's interpretation of the ROI review and also using suggested/standardized terminology. In one study, the implementation of HistoMapr-break performed well and showed 83% f-profile measured agreement of ductal epithelial dysplasia (N-300 WSIs and about 2000 ROIs).

FIG. 7 shows the micro-spatially distinct features in the Breast ROI identified by an embodiment of HistoMapr-Breast. HistoMapr identified four catheter ROIs, representing: (A) ductal epithelial atypical hyperplasia (ADH)702, (B) Flattened Epithelial Atypical (FEA)704, (C) Columnar Cell Change (CCC)706, and (D) normal ductus 708. HistoMapr-break provides visualization of the architectural patterns it finds in the ROI and overlays these patterns on the original image, as can be seen in ROI 702-. The pattern is derived from a combination of cytological and architectural features that are visualized as color-coded objects. A total of 18 different features are identified in each of the four ROIs and different colors are assigned to the various objects, as shown by the x-axis of the histogram 710.

Note the relatively high population of pattern #5 in ADH 702, pattern #7 in FEA 704, and pattern #15 in normal catheter ROI 708. As used herein, the term relatively high population of patterns in the ROI generally means that the pathologist or trained technician can see that this particular pattern occurs more frequently relative to the normal or baseline ROI. The relatively high population of patterns does not necessarily mean that the pattern appears more frequently than any other pattern in the ROI to be evaluated. Histogram 710 supports this observation, where we measure the relative proportion of the structural pattern in each of the following categories, respectively: ADH, FEA, CCC, and normal catheters. These architectural patterns are often named by our pathologist based on the visual characteristics they represent.

In addition to pathologist guidance, HistoMapr embodiments include features for additional applications/uses, such as truth data labeling, image content-based distribution of pathology, and archiving of critical diagnostic information for longitudinal continuity of care or clinical studies, as depicted in fig. 1A-1C. The truth data label is essential for machine learning training, but has historically been a bottleneck. Poorly implemented marking tools can also waste rare pathologist time.

Some embodiments of HistoMapr may effectively address this problem both with automated ROI discovery and with pathologist-friendly interfaces. As described above, embodiment HistoMapr may analyze one or more WSIs and extract ROIs from the respective WSIs. Using an efficient interface, a pathologist can quickly apply diagnostic labels to the identified ROIs. In some experiments, a pathologist was able to mark approximately 1000 ROIs identified by an embodiment of HistoMapr within an hour.

Fig. 8 depicts a pathologist-centric Graphical User Interface (GUI) for collecting annotations from a segmented region of interest. We employ the GUI in a Java environment to be easily used in any operating system, as the implementation of HistoMapr can be used in a pathologist's workstation or personal computer. However, the GUI may be implemented using any other programming language or scripting language. In some cases, the HistoMapr implementation is a plug-and-play system installed in an encrypted USB drive along with the ROI image to be marked. GUI designs for true value annotations are easy to learn and efficient to use. The pathologist does not need to manually draw or type in their input. Instead, they are shown with a series of ROIs, one ROI 802 at a time, and are asked to click on a display button in a keyboard or set of buttons (not shown) to classify the shown ROIs. In one example, the pathologist is asked to click on one of the buttons from 1 to 5, each corresponding to the labels "benign", "high risk benign", "DCIS", "invasive cancer" and "others".

The various ROIs 802 are shown with green segmentation boundaries 804 to collect feedback from the pathologist regarding the quality of the segmentation results produced by the HistoMapr embodiment. For example, the pathologist may use the option "other" in case they cannot decide on the nature of the shown ROI or determine that there is a problem with the identified segmentation boundary 804. The pathologist may be asked to complete the set of 250 ROIs at once, with the option of saving to be able to continue the procedure at a later time. A relational database can then be created to store, manage and retrieve the truth annotations, i.e., labeled ROIs. The labeled ROI may be used as a training data set to train embodiments of HistoMapr. During diagnostic use, correction markers derived during diagnostic work may also be included in the training dataset and used to improve HistoMapr based on real-world pathological diagnosis.

In some embodiments, the previously mentioned confidence scores are automatically generated by HistoMapr during WSI preview and analysis. The aggregate confidence score data associated with the different ROIs can then be used to estimate the difficulty of the case, and also combined with the number of ROIs to estimate the amount of work time required to observe the case. This allows for work diversion based on case attributes; difficult cases may be assigned to a specialist's second specialist pathologist rather than the general pathologist. The pathological work may also be evenly distributed to the pathologist's library, thereby increasing the efficiency of the pathologist's time usage.

xAI the derived statistics of the ROI or the entire case are fingerprints that can also be used for other content-based categories of purpose. The statistics include the features present, the quantification of these features, and/or the confidence scores. Typically, the confidence score is the confidence in the diagnostic label. Diagnostic labels are typically based on quantitative features, where quantitative scores for individual features may contribute to a confidence score. The confidence scores allow matching of the ROI to a library of known diagnostic images for decision support, education, quality assurance activities, or image content search and retrieval. Content-based analysis may also facilitate clinical trials by supplementing or automating patient screening and/or central review of pathology.

Finally, as described above in the xAI discussion of causal relationships, in some embodiments, archived HistoMapr information may be used to facilitate subsequent revisiting of previous biopsy cases. This includes comparing old biopsies to newer materials or centralized review clinical trials. The pathologist can easily observe the diagnostic ROI without the need to manually search for WSI. Previously recorded computational features may also play a role in the research to re-evaluate a diagnostic entity for risk or diagnostic criteria.

Some embodiments of HistoMapr are designed for transmitted light applications, where tissue samples are treated with hematoxylin and eosin (H)&E) Immunohistochemistry (IHC) labeling and/or other staining (e.g., special staining, chromogenic in situ hybridization, enzymatic metallography, etc.). Also provided in some embodiments is a companion platform for intelligent guidance of multiplexing to super-multiplexed fluorescence, referred to as TumorMapr^TM。

Summary of characteristics of interpretable AI/HistoMapr

In various embodiments, HistoMapr may effectively change a pathologist's view of a case from one or more WSIs to a guided series of shunted, diagnostically relevant regions of interest (ROIs). HistoMapr previews the entire WSI to discover relevant structures/features. For Breast core biopsies, HistoMapr-break uses Pairwise Mutual Information (PMI) based spatial statistics to discover vessel, vascular structure, and stromal features. By analyzing the PMI map, HistoMapr can find the ROI that contains these structures. In some cases, the ROI represents a single structure. For the purpose of speeding up the observation of easier/benign ROIs, the influential ROIs are typically smaller, have fewer structures, and fewer diagnostic ROIs (i.e., those at the end of the shunt list) may represent several benign ROIs merged together. Once the ROIs are identified, HistoMapr then analyzes certain of their characteristics. For example, the implementation of HistoMapr-break uses 18 features to classify the ROI as atypical or not (fig. 7). Each of these categories may include two or more sub-categories, and HistoMapr may generally classify the ROI into one or more sub-categories.

Typically, there will be several features in a set of features, and the machine learning process itself may determine that a subset of these features may reliably indicate a condition of interest. In one embodiment of HistoMapr, the AI finds 18 features to be significant, and 3 of them are the most informative, out of a total of 33 features considered. The other 15 features serve some purposes but are not as influential as the others. In this analysis, HistoMapr finds which features are present in the ROI, and it quantifies each feature it finds. The ROI analysis is a pattern that can be matched to the library of diagnostic tags of HistoMapr. The combination of diagnostic markers and ROI quantification is then used to shunt the ROI. Thus, for each ROI, a feature vector will be generated. For example, individual elements of the vector may include counts of particular features in the ROI. A machine-learned classification or prediction function will be applied to classify the ROI as, for example, ADH, FEA, CCC, and/or normal. In a boundary ROI, each ROI may be assigned only one classification, or in some cases, one or more ROIs may be assigned multiple classifications, each with a likelihood score, i.e., when the likelihood scores of two or more possible diagnoses are similar, HistoMapr may present a best-fit diagnosis as well as other diagnoses that are nearly possible. ROIs are to be shunted based on the severity of the category to which a particular ROI is determined to belong. For example, a ROI classified as ADH would be shown first to a pathologist.

For Breast core biopsies, HistoMapr-break can shunt the ROI from invasive cancer at one end to low risk benign at the other end in the clinical map (fig. 5). In an interactive work session, the pathologist reviews the entire case ROI by ROI in a split fashion. This approach is efficient because HistoMapr presents the pathologist first with the most clinically influential ROI; this guidance enables the pathologist to focus first on the most difficult decision. HistoMapr also continuously tracks ROIs that may require further disease examination with additional staining if necessary, or that may require consultation with another pathologist.

HistoMapr may also accelerate the labeling of truth data for training machine learning. The pathologist can quickly view the ROI and provide streamlined diagnostic labels. This unique approach can provide hundreds of labeled ROIs per hour. HistoMapr is much more efficient than traditional truth labeling in automated ROI discovery and effectively addresses bottlenecks in machine learning. In particular, the ROI discovery may be used to generate a labeling dataset used to train xAI the system. Also, based on what labels the pathologist has assigned, the xAI system can determine if the training data set is biased. HistoMapr has a pathologist-centric Graphical User Interface (GUI) for efficient annotation of segmented ROIs, employing a GUI that can be implemented on a pathologist's workstation or personal computer. The GUI design for true value annotation is easy to learn and efficient to use (fig. 8). Pathologists do not need to manually draw or type their input; instead, they are shown a series of ROIs and asked to select from one of the labels: "benign", "high risk benign", "DCIS", "invasive cancer" and "others". The ROI may be shown with a green (or another suitable color) segmentation boundary to collect feedback from the pathologist regarding the quality of the segmentation result. The pathologist may use the option "other" in case they cannot decide that the labeling of the ROI or that there is a problem with the segmentation boundary. The pathologist may be asked to complete a group of up to 250 ROIs (or another suitable number, e.g., 50, 100, 300, etc.) at a time, with the option of saving their progress to be able to continue at a later time. HistoMapr has a database to store, manage and retrieve truth annotations. In some experiments, the pathologist was able to mark about 1000 ROIs within one hour after being used on the platform.

Interpretable ai (xai): analysis of HistoMapr output typically produces a number of features (quantifies these features) that can be described in a pathologist-friendly language, linking the influencing features with diagnostic tags that can be used to interpret the HistoMapr tags in subsequent clinical sessions, which provides the pathologist with all the information they need to make the best possible decision. In one study, using a computational pathology pipeline (fig. 4), we extracted a breast ductal region of interest (ROI) in breast core biopsies and then shunted it as atypical contrasts not atypical (n-46 cases). To this end, 269 de-identified full-slice images (WSI) were generated (using 20-fold ScanScope XT of the leica biosystems, buffalo, illinois, usa). AI uses a combination of nuclear morphology and architectural placement to generate 18 micro-space catheter features that describe 95% of the training set (1009 ROIs) (fig. 7). The ROI was reviewed in the context of 18 features, particularly features 5, 9 and 15 that were seen differently in catheters whose atypical contrast was not atypical (as also seen in histogram 710 in fig. 7). Feature 5 appears to be related to the architectural rigidity and cobblestone-like appearance of the cells; feature 9 appears to represent the density of the matrix immediately surrounding the catheter; and feature 15 appears to be associated with the hyperplasias of the catheter. Using this information, an embodiment of HistoMapr incorporating xAI was created. This embodiment provides an AI feedback overlay, automatic presentation of feature examples, etc. (fig. 7).

xAI real-time visualization: "why? The "User Interface (UI) button may provide full transparency by presenting additional information in real-time that explains the tags of HistoMapr, enriches relevant differential diagnoses (where xAI determines that a particular ROI is ambiguous and assigns two or more tags thereto for the pathologist's ultimate decision), and confirms the strength of the HistoMapr analysis via one or more confidence scores. HistoMapr may display the results of his ROI analysis to the pathologist. The implementation of HistoMapr-break includes "why? "push buttons (fig. 6). HistoMapr finds features in the ROI and then quantifies them as previously described. When the pathologist presses "why? "button, she or he can see a visualization of ROI analysis by HistoMapr in a pathologist-friendly language (e.g., using terms such as strong rigid architecture, highly monomorphic nuclear patterns, etc.).

In various embodiments, the pathologist may click on "why? The "or" explain "button (fig. 6) accesses an interface panel or page to explain xAI the output. Fig. 9 illustrates an example display panel 900 displayed by an embodiment of HistoMapr. In some cases, a window or page may be displayed instead of a panel. The features detected in the ROI 902 are listed along with a quantification of one or more features. In some cases, a feature may be associated with a numerical or qualitative score, such as high, mild, moderate, and so forth. A confidence score analysis of the features (e.g., 0.5, as shown in fig. 9) may also be displayed on panel 900, which transparently provides an estimate of the intensity of its labeling and the difficulty of the ROI by HistoMapr. The confidence level may be determined based on the presence of interpretable features derived from the ROI. The user may forward the ROI to others, such as an expert, for consultation without ambiguity.

In various embodiments, the confidence score associated with the ROI is based on a quantitative synthesis of the individual features and/or further based on a labeling of the ROI. In some cases, examples of similar ROIs from other cases may be presented that may serve as reference guidance for decision support. For example, fig. 10 depicts a panel 1000 that includes a ROI 608 (fig. 6), which is labeled ADH. In response to clicking on "why? "button 604, showing information about the feature. In clicking on "why? "button 604, cover panel 1002 is shown. The cover panel 1002 includes two reference ROIs 1004, 1006 from one or more other cases, where the reference ROIs are also labeled ADH and have similar features to those shown in the cover panel 602 for the ROI 608 being reviewed. A pathologist may evaluate the similarity of different ROIs in determining the correctness of the labeling of the ROIs 608. By visually comparing the ROI 608 being reviewed with other ROIs having the same label (e.g., ROIs 1004, 1006), the reviewer may feel more confident about the specified label in which such labeling is accepted.

In some implementations, the schematic representation of the feature in question has controls (e.g., sliders) that allow the pathologist to observe the low-to-high continuity of the feature, with the features being sorted based on their importance to the particular situation marked for the ROI. In some embodiments, the HistoMapr xAI system presents its differential diagnosis and displays the pros and cons of the various diagnoses considered. If ambiguous, the HistoMapr embodiment may suggest further disease examination with staining or expedite (e.g., online) consultation with another pathologist.

xAI some embodiments of the system provide for patient biopsy shunting based on confidence score (content based). Cases above a certain difficulty threshold may be assigned to a secondary professional expert pathologist, while other cases may be sent to the front-line general pathologist. The estimates of case difficulty provided by the HistoMapr embodiment and the image volume may be used to more evenly distribute pathological cases to a group of pathologists in order to better utilize professional resources. In particular, as described above, embodiments of HistoMapr provide data in support of their diagnostic markers, and may include a measure of the strength of the decision (e.g., confidence score) of HistoMapr, which may indicate whether the decision is simple or ambiguous/difficult.

This is a strong communication from HistoMapr to the pathologist, as it allows the pathologist to understand why HistoMapr marks the ROI as it does and whether HistoMapr considers the ROI difficult or ambiguous. This allows the pathologist to have all the necessary information for making a diagnosis, and it allows the pathologist to fully examine the performance of HistoMapr in real time. At a higher level, this assessment can also be exploited to drive case diversion; difficult to visualize cases may be routed to a secondary professional pathologist for review rather than the first line of ordinary pathologists.

Fig. 11 illustrates an example offload process 1100. At step 1102, an embodiment of HistoMapr previews and analyzes cases (e.g., one or more WSIs), identifies one or more ROIs, and assigns labels and confidence scores to the respective ROIs. At step 1104, it is determined whether one or more ROIs indicate critical discovery based on the specified label. For example, a marked invasive carcinoma or Ductal Carcinoma In Situ (DCIS) may be considered severe, whereas a marked low-risk benign lesion may not be considered severe. If one or more markers of the case ROI are determined to be critical, the case may be forwarded to an emergency pathologist (step 1106).

The labeling of the various ROIs of the case is not determined to be severe, and the confidence scores of the different ROIs can be considered in step 1108. If the confidence score or one or more ROIs (up to a threshold number of ROIs) is less than a specified confidence threshold (e.g., 40%, 50%, 75%, etc.), the case can be forwarded to a secondary specialist (e.g., a cancer secondary specialist, an ADH secondary specialist, etc.) selected based on the label (step 1110). The threshold number of ROIs may be a specified number, e.g., 3, 5, 10, etc., or it may be a percentage of the total number of ROIs associated with a case, e.g., 10%, 20%, 50%, etc. In some cases, the confidence scores of the different ROIs may be aggregated to derive a confidence level for the case/WSI, such as by selecting a maximum of the confidence scores for the different ROIs, based on a simple average or weighted average of the size of the ROIs and/or the severity of the label assigned to the ROIs, or the like.

If the confidence score or confidence level of the ROI is determined to be acceptable in step 1108, then a workload estimation score may be calculated based on the number of WSIs analyzed and/or the total number of ROIs detected in step 1112. In step 1114, using the workload estimation scores, cases may be evenly assigned to the pathologist library. Thus, in some embodiments, HistoMapr may shunt patient biopsies based on the confidence level in the case and key findings. Cases above a certain difficulty threshold may be assigned to a secondary professional expert pathologist, while other cases may be sent to the front-line general pathologist.

Using standardized terminology for diagnostic tags, embodiment HistoMapr may encourage pathologists to report results in a more uniform manner, thereby improving the quality of the report. When the pathologist and HistoMapr together complete a case, HistoMapr can use the pathologist's decisions to build a pathology report in real time using a standardized language. This may standardize pathology practices and, due to standardization, may improve the quality of practice reports by making them more understandable. Examples of such standardized reports are shown in fig. 1A to 1C.

Typically, cases (e.g., WSI) or their individual ROIs are annotated by an embodiment of HistoMapr via xAI. In addition to diagnostic tags, there are features, feature quantitative (numerical and/or qualitative), and confidence scores. These form fingerprints that can be used to match the ROI with other cases, with a library of known diagnoses, or with cases of other pathologists for a variety of purposes (fig. 1). Some uses of annotated WSI/ROIs include: (1) decision support-examples of similar ROIs from other cases, from a library of known diagnoses, or from instructional material can be shown to a pathologist to assist in making decisions regarding the analyzed ROIs. (2) Educational-comparison of ROIs to known examples of good diagnosis, and comparison of ROIs to nearly similar ROIs with different diagnoses (i.e., presenting differential diagnosis). (3) Quality Assurance (QA) activities-partial automation by other pathologists with a second review of the pathologist's work (e.g., QA review, standard second opinion scenario, etc.). Moreover, continuous monitoring of HistoMapr performance is part of the QA framework of pathology practice. (4) Clinical trials-computer-aided screening of patients' applicability to potential clinical trials streamlines central review of pathological material (i.e., ROI presentation, without having to review entire cases). (5) Content-based image searching for the above purposes or for other reasons.

Importance and evaluation of HistoMapr xAI System

The HistoMapr embodiment may use xAI to assist pathologists in a transparent manner and may guide them in making the best diagnosis they can make. This is a unique aspect of HistoMapr, since AI techniques currently known for pathology are based on deep learning. It is currently not possible to ask the deep learning system "why? ", and it is not possible to obtain insight into the analysis performed by the AI. However, interpretable AI is an important feature of early computer pathology systems, as it can help pioneer pathologists start building trust even with highly validated intelligent software guidance, as schematically shown in fig. 12.

Another important aspect of using the xAI technique comes from future possible AI usage regulations in the decision making process that may affect human rights and/or health. Regulations set forth before the european union prohibit "automatic processing" unless the user's rights are protected. The user now has "explanatory power" on decisions created by the personal information based algorithm/machine. Future laws may further limit the use of AI in professional practice, which may present a significant challenge to the industry. Moreover, business and clinical owners may not target the full automation of machine learning models, as they may not blindly trust the models, as there is no quantifiable confidence in how conventional machine learning models work, and the transparency of decisions of conventional AI is often insufficient. Regulatory constraints in which the healthcare must operate provide opportunities for interpretable AI, and the HistoMapr xAI system can introduce a level of trust to the pathologist using an interpretation interface.

In the design of the HistoMapr xAI system, key concepts for evaluating xAI system measurements and human-machine performance were considered. These key concepts include goodness of interpretation, satisfaction of interpretation, desirability of learning metrics, and user performance. In short, the goodness of interpretation represents the clarity and accuracy of the interpretation provided by the xAI system. The evaluation of the criterion can be performed by: the user is simply asked a yes/no question to learn: a) interpreting whether to help them understand how the software works, b) interpreting whether to be satisfactory, detailed and complete, c) interpreting whether to be executable (i.e. to help the user know how accurate and reliable the software is), d) interpreting whether to let the user know how trustworthy and reliable the software is, and e) interpreting whether to let the user know that the software is multi-valued. In various embodiments, HistoMapr provides detailed feature interpretations (e.g., confidence levels, impact of features on decisions, detailed interpretation of features, let users take action on output) in the interpretation interface, from which it can be seen that the goodness of interpretation metric for HistoMapr may be higher.

The interpretation satisfaction is to inform the user of the degree to which they feel they understand the AI system or the process they interpret. This criterion is common to many xAI systems because it is intended to meet the goals of the user in understanding the AI process. To meet this criteria, the xAI system should be able to solve the user's problems, such as, a) how do i avoid failure mode? (showing the user's desire to mitigate errors), b) what do i do if you go wrong? (showing the user's desire to avoid errors), c) what is it implemented? (showing the user's interest in understanding the system's functionality), d) what did it just do? (a sense of satisfaction of the user with an understanding of how the system is being implemented to make certain decisions), and e) why does it not do "z" but rather "x"? (showing the user's solution to the awareness that an understanding of the decision is being implemented). The HistoMapr xAI system may solve the interpretation satisfaction criteria by answering the user's questions via its interpretation interface. Why in HistoMapr GUI? The "button is specifically designed to address these issues by explaining why HistoMapr believes that the ROI belongs to a particular type of lesion.

Knowing the desire metric is also an important criterion, as interpretation can suppress knowing the desire and reinforce the flawed artifacts. This can occur in a number of ways: a) explanations may be overwhelmed by details, b) the xAI system may not allow or may make it difficult for users to ask questions, c) explanations may be silent because they lack knowledge, and d) explanations may include too many open variables and fragmentary material and require less awareness as confusion and complexity increase. For these reasons, the evaluation of the perception of interest by the user may be informative in the evaluation of the xAI system. The HistoMapr xAI system is designed to receive feedback from end users by asking simple survey questions. Using user feedback, the information panel can be customized to achieve a high learned desirability metric.

Finally, the user performance, including the metrics of the joint user system performance, will improve as a result of being given a satisfactory interpretation. The main aspect of the user performance measure is the user's performance quality, such as the correctness of the user's prediction of what the AI will do. For these aspects of performance, just like HistoMapr's performance, we measure response speed and correctness (hits, errors, misses, false alarms), but in this case, the user's prediction of machine output is measured. Both typical and atypical cases/situations can be examined. In addition, HistoMapr can measure the correctness and completeness of user interpretation of machine output for rare, abnormal, or anomalous cases.

HistoMapr-Breast: validation study

The main objectives of this study were: i) validating embodiments of HistoMapr-Breast (HMB) on a large number of breast biopsy Whole Slice Images (WSIs); ii) collecting truth annotations from expert pathologists about the ROI, wherein such annotations may be used to train embodiments of HMB; iii) detect high risk benign breast lesions with high accuracy and high efficiency, which are very challenging for pathologists to reach consensus with high consistency; and iv) demonstrate that embodiments of HMB can increase the efficiency and accuracy of pathologist "calls" and can also increase consistency among pathologists.

For these goals, we a) collected 4865 breast biopsies WSI; b) testing and evaluating the performance of different embodiments of HMB in differentiating broad spectrum breast lesions (benign, atypical hyperplasia ("atypical"), Ductal Carcinoma In Situ (DCIS) and invasive carcinoma); c) demonstrate WSI navigation based on diagnostically shunted (benign to malignant) ROIs; and d) testing the consistency between the performance of 3 expert pathologists and 2 training pathologists, both with and without the help of HMB.

A total of 2171 cases selected by the pathologist included 3347 breast core samples. A total of 4865 tissue slides from the 3347 samples were imaged and de-identified using an Aperio ScanScope AT2 AT 0.5 micron/pixel resolution (20 x magnification). The resulting 4865 WSIs processed by an embodiment of HMB to segment about 201000 ROIs. We selected 4 major diagnostic categories (invasive cancer, DCIS, high risk tissue and low risk/benign tissue) to map 24 different diagnoses for potential annotation as shown in table 1 in fig. 13. Although table 1 shows several subcategories within each main diagnostic category, in our experiments we used only four main diagnostic categories for truth labeling. In other cases, one or more subcategories from one or more major categories may be used for truth tagging.

The truth data label is essential for machine learning training, but has historically been a bottleneck. Poorly implemented marking tools also waste little pathologist time. Embodiments of HMB can effectively address this problem with both automated ROI discovery and pathologist-friendly truth annotation GUI.

Various embodiments of the HMB feature an annotation tool with a Java-based GUI, and these embodiments or annotation tools can be easily installed on many operating system platforms. FIG. 14 depicts one embodiment of a quick annotation GUI that allows a pathologist to navigate through a ROI detected by HistoMapr and mark the ROI using keyboard shortcuts. The field of view in the interface is generally centered on the ROI 1402. Side panel 1404 shows available categories 1406 and corresponding keyboard shortcuts.

Unlike many other annotation systems, the GUI in embodiments of the HMB does not require the pathologist to manually draw or print/type their input. Instead, in one embodiment, they are shown with a series of ROIs one by one and are asked to click on the keyboard buttons from 1 to 6, 1 to 6 corresponding to the labels "wettability", "DCIS", "high risk", "low risk/benign", "unknown" and "other/image problems", respectively. In some embodiments, the buttons may be provided on the screen as part of the display. Depending on the number of classes or categories, fewer or more than six buttons are contemplated.

In the experiment, first, using the GUI, 4462 ROIs were required to be marked by each of three pathologists. The embodiment of HMB used does not provide any guidance at this stage, i.e. the HMB does not have a label suggesting the ROI. At least two ROIs in three pathology families agreed (majority vote) on the label were named consensus marker set, which includes 4281 ROIs. The pathologist fully agreed to the diagnostic label of 3172 ROIs, which resulted in an overall agreement of 71%. From the 4281 consensus labeled ROIs, we selected a 1077 ROI test image set that was balanced between the four major diagnostic categories. For consistency analysis, the Fleiss' kappa for this test set was calculated as 0.66. We also evaluated the pathological manifestations of two clinical instructors ("breast/gynecologic colleagues") using 1077 ROIs. Based on the consistency labels from three experts, their accuracy was about 67% in the four main categories, as shown in table 2. In Table 2, "without using HistoMapr-Breast" means without using guidance from HMB; a GUI is used.

We then rerun the experiment, but this time the embodiment of HMB used was configured to provide guidance by assigning respective xAI-generated labels to the respective ROIs. FIG. 15 depicts an example GUI 1500. The ROI 1502 is shown, and on the right, in panel 1504, a recommended or suggested label 1506 of HMB is shown. Key findings 1508 (i.e., features that explain the tags) are also shown in the right panel 1504 along with categories 1510 available for tagging to assist the user during annotation and/or diagnosis. In this stage of the experiment, we tested the effectiveness of embodiments of HMBs by showing the pathologist the ROIs and the recommended labels from the HMBs for a given ROI to determine if their performance improved.

Specifically, three expert pathologists were asked to mark 1077 selected ROIs again, but this time using the guidance of the HMB suggested label, as shown in fig. 15. Their consistency improved as indicated by an increase in Fleiss' kappa to κ of 0.75. We also re-evaluated the performance of both colleagues using 1077 ROIs, but using the guidance of the HMB suggested label. Their performance was significantly improved, as shown in table 2, with an increase in accuracy from 67% (no guidance) to 91% (with guidance). In Table 2, "use of HistoMapr-Breast" means use of guidance from HMB and GUI. Thus, the embodiment of HMB used demonstrates increased diagnostic accuracy and performance of colleagues. It follows that embodiments of the HistoMapr platform can greatly enhance annotation of images of digital pathology.

TABLE 2-Performance of clinical colleagues in diagnosing 1077 ROIs

Experiment: embodiments using HMB differentiate breast pathology profiles from benign to atypical to DOS to invasive cancer

As described above, the HMB computing pipeline typically includes WSI stain color normalization, catheter ROI segmentation, and cytological/architectural feature extraction to further classify the catheter ROI. The process may analyze large WSIs at a rate of about 90 seconds per WSI. With expert management, we retain a balanced distribution of breast lesion classes in the image dataset. We trained our AI system with 50% of the annotation data and used the remaining annotation data for validation and testing.

For the cytological phenotype of the ROI, we generated an accurate set of nuclear masks for each ductal ROI. Each nucleus is characterized by 196 features, a packageIncluding morphological, intensity, and texture features. As a result of nuclei that are normal, atypical and polymorphic in high-risk benign breast lesions, we found three dominant phenotypes (nucleoli)₁、nuclei₂、nuclei₃). For architectural phenotype analysis, the organization is represented by 5 different objects: three nuclei (nucleoli) for cytological phenotypic analysis₁、nuclei₂、nuclei₃) And two superpixel-based components (matrix and lumen). A spatial network is constructed by breadth-first traversal from the individual objects. Neighborhood statistics for each object were collected and clustered into q different architectural patterns using k-means, which covered 95% of the input variance.

We construct the architectural feature vectors for three different scenarios based on: (i) color-based architectural features (AF-C) using superpixel derived nuclei (i.e., stroma and lumen objects); (ii) architectural features of the nucleus (AF-N) based on cytological phenotypic analysis, using the nuclear phenotype alone; and (iii) combined architectural features (AF-CN) that use nuclear phenotypes in combination with matrix and luminal superpixels. Without a defined conduit structure, such as in invasive cancer, we expect that the cytological phenotypic analysis that captures the frequency and co-occurrence of the nuclear phenotype is sufficient for classification. More complex features, such as our architectural phenotype, can be considered a more refined sub-classification for infiltrating ductal lesions. The runtime performance of the implementation of HMB for analyzing WSIs averages less than 2 minutes per WSI on an 8-core 64 GB RAM workstation.

We have discovered additional distinguishing architectural phenotypes to extend our vocabulary of features for xAI. Fig. 16 depicts these distinguishing architectural phenotypes. Panel (a) shows the relative proportion of the most distinctive architectural phenotypes in each of the following categories (8 out of 67 analyzed): invasive cancer, DCIS, high risk lesions and low risk/benign lesions. Note the following relatively high population of architectural phenotypes: #1 in invasive carcinomas and DCIS; #2 in high risk; and #3 in low risk/benign lesions. All of these phenotypes have been further reviewed by expert pathologists as they are shown with the highest proportion of ROIs containing the phenotype. Panels (B) - (I) in fig. 16 depict ROIs comprising dominant phenotypes #1 to #8, respectively. These phenotypes have the following histological interpretation:

phenotype #1 (panel B) is a "nuclear texture smoothness" feature; phenotype #2 (panel C) is a "uniform spacing between nuclei" feature; phenotype #3 (panel D) is a "benign stroma" feature; phenotype #4 (panel E) is a "benign stroma of cell filling" feature; phenotype #5 (panel F) is a "apocrine and atypical nucleus" feature; phenotype #6 (panel G) is an "atypical & malignant nucleus" feature; phenotype #7 (panel H) is a "variable inter-nuclear distance" feature; and phenotype #8 (panel I) is a "larger nucleus" feature. These interpretations are later used in some embodiments of the HMB to provide an interpretation to the user, such as in panel 602 shown in fig. 6.

Following discussion with expert pathologists and practicing clinicians, we devised a hierarchical classification strategy to suit the purposes of case diversion and retrospective QA, where first an invasive versus non-invasive invocation is made, then the non-invasive diagnosis is classified as either pre-invasive lesions (atypical and DCIS) or benign, and finally the DCIS or atypical classification steps are applied. Invasive cancer destroys normal tissue structures, so it is placed higher up on the classification ladder to customize the feature space for more difficult high risk atypias.

Since cytological and architectural phenotypic analysis relies heavily on accurate vessel segmentation, we tested implementations of the HMB pipeline by using an improved vessel segmentation algorithm (described in appendix a) for the most challenging of all tasks: low risk (normal tissue and Columnar Cell Change (CCC)/Columnar Cell Hyperplasia (CCH)) versus high risk (flattened epithelial atypical (FEA) and ductal epithelial atypical hyperplasia (ADH)) benign lesions.

The cross-validation experiment was repeated 100 times, and the sub-sampling step was used to separate training and testing datasets before reporting the average accuracy, which ensured that the training data was approximately equal in all diagnostic categories. We tested different embodiments of the HMB pipeline with na iotave bayes, decision trees, random forests, Support Vector Machines (SVMs), and logistic regression, as well as artificial neural networks used as classifiers. The performance of our model using color architectural features (AF-C), nuclear architectural features (AF-N), combined architectural features (AF-CN), and Cytological Features (CF) was observed.

Table 3 summarizes the classification performance results for the three cohort categories: wettability versus non-wettability; DCIS contrasts high-risk and low-risk tissues; and high risk tissue versus low risk tissue. As can be seen in table 3, the decision tree using AF-CN yields the best performance on a test set of 1077 ROIs. In some embodiments of HMB, the spatial organization and morphometric properties of the nuclei and the lumen may be taken into account to further enhance the classification performance score. Although there are several studies on cancer detection in breast tissue images, to our knowledge, our studies using embodiments of HMB are the first of its most challenging tasks to solve the classification of high-risk versus low-risk benign breast lesions from WSI with high accuracy.

Table 2-evaluation of the performance of the embodiment HistoMapr-Breast on three Breast core biopsy cohorts

Intelligent case diversion

As described above, one benefit of the HistoMapr xAI platform of embodiments is the use of the xAI feature to stratify all slices belonging to a patient (e.g., breast core slices, brain slices, etc.). For example, embodiment HistoMapr may analyze all WSI of a patient biopsy by locating and segmenting the ROI. These ROIs can be further ranked according to their diagnostic importance based on classification results. For example, an embodiment HistoMapr-break (hmb) may order the ROIs as: malignant, then DCIS, then atypical, and finally benign. Some embodiments of HistoMapr use this information to provide a three-stage shunt system that can prepare and shunt all cases in pathological practice. The three levels of shunting are described below.

ROI splitting: and finding, segmenting, spatially characterizing and shunting the region of interest according to the severity and complexity of the region of interest. This may assist the pathologist in first assessing critical areas on the diagnosis and may also assist in making the diagnosis. In addition, the observation xAI feedback can help pathologists make critical diagnostic decisions very early in the review process. After reviewing the critical ROIs first, the pathologist can quickly go through the remaining ROIs, as decisions about the case may have been made by the pathologist looking at the most critical ROIs first. Pathologists may also access the WSI from which the ROI was identified, where they may visually scan the rest of the tissue (i.e., the non-ROI areas) for final diagnosis.

Fig. 17 distinguishes between the current review and analysis process without the HistoMapr platform ("non-HistoMapr process") and the enhancement process with the implementation of the HistoMapr platform ("HistoMapr process"). In particular, in the non-HistoMapr process, the reviewer must go through each WSI in sequence, each time analyzing a selected region. The reviewer has no prior knowledge as to whether the region to be observed is critical to the diagnosis. This process is prone to errors since some regions may be missed. Also, critical areas may be examined sooner or later based on the scan order. Furthermore, ambiguous areas that are difficult to decide can be observed/examined before relatively unambiguous areas that are relatively easy to decide. Strict scan sequences typically do not allow a reviewer to benefit from analysis of easy/unambiguous regions in the analysis of difficult/ambiguous regions unless the reviewer returns to the region.

Embodiments of the HistoMapr procedure can avoid many of these problems because the embodiments of HistoMapr used first identify and separate important or assay-related regions of WSI from regions that are not important or relevant to the assay. These relevant, important, critical areas are clearly marked with appropriate boundaries (e.g., as shown in fig. 2B, 5, 6). Furthermore, embodiments of HistoMapr provide suggested diagnostic tags and, upon user request, provide an explanation of how the region was identified as a ROI, how the suggested tags were derived, and a confidence score indicating the confidence of the platform in its determination. (see fig. 6 and 9 and the accompanying discussion). Based on the severity and/or confidence score of the suggested tags, the ROIs deemed most critical are presented to the reviewer first, followed by ROIs deemed less critical. This process may ensure that the entire WSI is reviewed without missing any regions, critical diagnosis may be performed early, which minimizes the delay in the discovery of critical conditions, and diagnostic decisions about relatively easy/unambiguous ROIs may inform diagnostic decisions about relatively difficult/ambiguous ROIs.

Slicing and shunting: in this case, the WSIs to be viewed/examined by the pathologist are sorted. To this end, embodiments of HistoMapr analyze each WSI in the set of WSIs, as described above. The WSIs may then be ranked based on the number of ROIs, the number of conditions identified, the severity of the ROIs (which may be the maximum severity, average severity, weighted average severity, etc. of all ROIs in the WSI), the confidence scores of the ROIs (which may be the maximum confidence scores, average confidence scores, weighted average confidence scores, etc. of all ROIs in the WSI), or a combination of two or more of these factors. The triage of WSIs, i.e., WSIs that are considered more critical (e.g., they may have a large number of ROIs, may have several conditions, may represent a severe condition, and/or may be less confident to the platform for analysis thereof) may be presented to a reviewer for analysis/review prior to other WSIs. It can be seen that as can be seen in fig. 17, the reviewer (e.g., pathologist) does not waste time looking first at WSIs that may not contain a diagnostic region.

Case diversion: in our discussion with pathologists, we understand that a large number of patients suffer from unbalanced case distribution due to random forwarding of cases to pathologists. To address this problem, some embodiments of HistoMapr may shunt cases and forward the cases to the appropriate pathologist according to their secondary expertise and/or workload. Fig. 18 shows a comparison of the non-HistoMapr procedure and the HistoMapr procedure. In the non-HistoMapr procedure, the case distribution system (or individual) has no prior knowledge about the nature of the case. It follows that if a difficult or tricky case is assigned to a general physician or a specialist not specifically addressing the problem of diagnosis of the case, there will be a delay in diagnosis as the case may need to be forwarded to the correct specialist/sub-specialist, as shown in fig. 18. This can also result in an overburdened general practitioner. Moreover, there is a risk of misdiagnosis. On the other hand, if relatively easy cases are forwarded to the specialist, this may result in overburdening the specialist and will take time for them to analyze more difficult cases. However, in the HistoMapr procedure, simple benign and malignant cases may be forwarded to the average physician, while borderline and difficult cases may be forwarded to the secondary specialist. This can help maintain case load balance and better use of time and expertise of specialist and sub-specialist. An embodiment of the HistoMapr process is described above with reference to fig. 11.

To apply xAI in the case diversion interface, some embodiments use the disease example classification and xAI according to "case severity," "case difficulty," and "atypical findings" for efficient diversion. For example, "severity" may be based on diagnosis, since cancer is severe and benign is not. In addition, if the embodiment HistoMapr is confident in its decision (e.g., the case has simple DCIS or invasive lesions), a non-specialist pathologist may be assigned to the case and the case may be designated to log out before benign cases. If the confidence of the HistoMapr embodiment is low, the case may be marked as "difficult" to analyze and it may be forwarded to the second specialist (e.g., if the HistoMapr embodiment is 70% confident that the case is DCIS, it will be designated to pass to the second specialist). Finally, if an embodiment of HistoMapr is confident that a case has "atypical findings," the case may be assigned to a secondary specialist. Thus, embodiments of case diversion may improve the throughput of pathological practices, e.g., by 25% or more. According to our simulations, the computational pathology using the various embodiments described herein can be up to 56% more efficient than traditional microscopy (e.g., for breast core biopsy).

We discuss above with reference to fig. 6 and 9, for example, valuable explanations that various embodiments of the HistoMapr platform provide regarding the marking decisions it has made. In a similar manner, some embodiments provide a reason or reason to explain the diversion decision. To this end, some embodiments use updated terminology in response to a pathologist using "why? The "or" explain "option to provide the answer. In particular, in some embodiments, for case diversion, each diverted case is associated with information explaining why the case was diverted in that particular manner. Fig. 19 depicts an embodiment of a GUI 1900 for case diversion. GUI 1900 includes a case/patient list 1902, a patient type indicator 1904 (e.g., indicating whether the patient is new or returned), and an indicator 1906 that describes the conclusions drawn by an embodiment of the HistoMapr platform for analyzing cases. For each case/patient, the GUI includes "why? "," explain, "or similar buttons 1908, and an indicator 1910 indicating to whom the case should be forwarded for auditing. The conclusion indicated by indicator 1906 may include "emergency findings," indicating that the case/patient may have a malignancy or another severe condition, and should be seen before other cases. Emergency pathologists are typically non-specialist anatomical pathologists who are not trained for any particular type of disease/anatomy, but they can easily and quickly diagnose malignant tumors and benign cases across all organs. Another example of indicator 1906 is "need to care," indicating that the HistoMapr platform is not highly certain (e.g., more than 60%, 80%, 90%, etc.) with respect to its conclusion, "high risk," indicating that a severe condition is detected with respect to a case, etc.

In clicking on "why? "button 1908, panel 1912 (or window or page) is displayed in which the HistoMapr platform will display information 1914 associated with the case and that can explain the particular triage of the case. One or more related ROIs 1916, which may further inform interpretation, may also be displayed in the panel 1912. These explanations/reasons may indicate the difficulty of a case (e.g., "difficult cases involving atypical findings," "easy cases with definite invasive findings," etc.) and why it is difficult or easy based on evidence quantified by the HistoMapr platform. Some embodiments of the HistoMapr platform also use a "shunt confidence score" to indicate its confidence in the shunt decision, which incorporates features, feature quantities, confidence in ROI classification, etc. within the case.

Fig. 20 schematically depicts a typical operating environment for an embodiment of the HistoMapr shunt system. The environment may include the following three modules: (1) a data center interface: an interface capable of handling access (e.g., security, VPN access) to WSI data stored at a remote data center (e.g., in pathology practice); (2) by a third party (e.g., Splntellx as shown in FIG. 20)^TM) Implementation of HistoMapr of operation. The data center interface will allow the HistoMapr embodiment to access and analyze WSI data and shunt ROIs, WSIs, and/or cases as described above; and (3) an embodiment of the HistoMapr case diversion tool installed on a computer of a diversion administrator or pathologist, which provides a GUI as described above.

WSIs can be generated, saved, filtered, and de-identified at third party partner facilities (e.g., pathologist laboratories/offices), and will be available to access implementations of the HistoMapr platform via the data center interface. The WSI may be processed in a chronological order by running xAI a core algorithm to detect suspicious conditions and then tagged based on the results of the analysis and the confidence scores. The WSI may be forwarded to a case diversion tool, which may access the analysis results of an embodiment of HistoMapr for performing the analysis and associated explanatory information from an embodiment of HistoMapr. The case diversion tool can also access patient data from the information system of the pathology practice. The case diversion tool can then automatically (or used by an administrator) perform diversion and prioritization, as described above. In doing so, case diversion can help distribute cases in a balanced and accurate manner by distributing critical cases to secondary or specialist physicians. In some cases, embodiments of the HistoMapr platform may also be installed in the pathologist's office, and the diversion tool may be an integrated component of the platform.

The case diversion GUI (e.g., GUI 1900 in fig. 19) can display a list of cases diverted from high priority cases (e.g., malignant, atypical) with a pop-up text notification of new cases with suspicious findings when those cases are ready. In the GUI, each case in the list may have "why? "or similar buttons. Click on "why of the case in the list? The "button pops up a textual explanation of the key findings from the case and a compressed, small, colored, unmarked ROI image entitled" non-diagnostic use "and displayed as a preview of the key findings, as can be seen in fig. 19. The condensed preview is used for informational purposes only and may or may not contain any comments/interpretations of the discovery. Presenting a list of triage cases and interpretation feedback to a pathologist may facilitate early triage by prompting the user to evaluate relevant original slices or WSIs. Thus, suspicious cases may be attended earlier than they would according to a procedure that does not use case diversion as described herein. Embodiments of the HistoMapr platform do not alter the original medical image. They can provide recommendations of who to assign cases based on analysis of the platform. The case diversion tool can be used under the control of a supervisor, administrator, or lead pathologist who will forward the diverted cases to the pathologist. Alternatively, the diversion tool may operate autonomously and may forward the case to the pathologist himself.

Review quality assurance

One of the most urgent unmet clinical needs for pathological practice is the tools and techniques that can reduce diagnostic inconsistencies among pathologists. Many studies involving diagnostic challenges report only moderate agreement among pathologists. For example, benign breast lesions are a significant source of inconsistency and uncertainty, which leads to inconsistencies of up to 52% in ADH diagnosis. For stubborn cases, such as atypical, the diagnostic accuracy often drops significantly if the pathologist is not a second specialist. These inaccuracies may lead to unnecessary surgical resection or to missed diagnosis of malignant tumors, which leads to an increased likelihood of litigation as a medical accident. ADFI diagnosis must be reliable because it presents significant consequences to patients and can be an economic burden to healthcare delivery systems due to the possibility of unnecessary surgery and the need for frequent screening by these patients. Thus, there is a significant unmet clinical need for techniques and tools that allow pathology practices to assess how consistent their pathologists are when calling and enable them to learn from poor consistency. Although some Quality Assurance (QA) mechanisms are known, such as randomly selecting 10% of the most recently reviewed cases for panel review, etc., they are generally neither effective nor efficient QA processes, as many of them are crude and semi-quantitative.

To address these challenges, embodiments of retrospective quality assurance tool RetroQA are provided. RetroQA may be part of the HistoMapr platform. In various embodiments, RetroQA may review previously completed cases in order to continue post-diagnostic monitoring. The xAI features of the tool can be used to help identify potential pathological outcome differences or errors. This important QA activity can reduce risk, increase clinician confidence, and help pathologists monitor their diagnostic work in a timely manner. As with other HistoMapr xAI applications, "why? The "button provides details identifying the QA problem (such as" the likelihood of malignancy of a biopsy diagnosed as benign "is high) and supporting evidence that may support the analysis performed by the tool, such as an ROI image with highlighted features (such as shown in fig. 7). In some cases, confidence scores are calculated in a similar manner for completed diagnoses, with confidence scores calculated for markers determined for embodiments of the HistoMapr platform. If the confidence score is low, then a user click "why? "button to provide message" very weak confidence score that caused diagnosis ". The RetroQA embodiment may improve the post-diagnosis accuracy of a pathologist by up to 25% in many cases, and even higher in other cases. According to our study (see table 2 above), the performance of the non-specialist in sensitivity to the truth of high-risk tissue diagnosis is moderate (only 27.9%). This performance can be improved with the aid of the RetroQA embodiment.

From our discussion with pathologists, we understand that many practices currently apply quality control to about 10-15% of cases, randomly selected from all cases analyzed over a particular time period. Cases were selected either retrospectively after completion or during ongoing analysis prior to completion for a second review. These cases were reviewed by a pathologist different from the one who analyzed the cases for the first time. Since the selection process is random, there is still a significant risk of missing misdiagnosed cases, which is especially critical for relatively difficult cases, such as atypical related breast lesions. Embodiments of RetroQA may also improve the selection process.

In various embodiments, RetroQA selects a case that can benefit from a second audit. Cases may be selected by analyzing WSIs associated with the case using an embodiment of the HistoMapr platform. The platform may specify a label to the ROI in the WSI and select the WSI if the assigned label is associated with a severe condition. Alternatively, or in addition, based on the assigned tags and/or confidence scores computed by the platform when assigning the tags, the platform may determine whether the WSI includes one or more difficult-to-decide ROIs. In a retrospective selection, the platform may calculate confidence score(s) for ROI diagnosis(s) in WSI. Based on any of the factors or scores described above, or a combination thereof, embodiments of RetroQA may determine that the case/WSI may benefit from further review and select the case accordingly.

The implementation of RetroQA is also "why? "," explain, "or similar button, which the pathologist performing the second audit can use to understand why the particular case was selected, for example, for a QA assessment. In clicking on "why? "buttons, a panel, overlay or page may be displayed, where RetroQA will present the factors and/or scores used in the selection process. Examples of explanatory information or reasons include an indication that a case may be difficult (e.g., a message: "difficult cases involving atypical findings, potential risk of misdiagnosis"), or other reasons (such as: "potential difference, confidence level 85%"). Additionally, in some embodiments, RetroQA may also provide supporting evidence of its selection, such as an ROI image with highlighted features (such as shown in fig. 7), which may support analysis of the tool.

Thus, embodiments of HistoMapr with RetroQA may analyze all previously completed cases for post-diagnostic monitoring, and may be used to identify cases to be forwarded for a second review (e.g., for quality control). Cases may be selected because HistoMapr/RetroQA may determine that these cases may have potential differences or errors with respect to the completed pathology results, and/or their analysis using HistoMapr results in a low confidence score, reflecting that the case is difficult to diagnose and potentially at risk of misdiagnosis. All cases selected by HistoMapr will be forwarded to the consensus of the pathologist in the pathology practice for a second review. This can help reduce risk by detecting significant differences (e.g., benign versus malignant), provide the supervisor with evidence of ongoing and effective QA activity, and improve clinician and patient confidence in the diagnosis, especially difficult diagnoses such as breast atypia. "why? The "button may provide transparency and show why there are potential discrepancies. This may help improve accuracy of practice and patient safety.

HistoMapr-Lung: analysis of lung tissue

Tissue staining has been widely used in the pathology of disease or subtype diagnosis, but it has limitations. Computational pathology, i.e., the processing and analysis of medical images using computational methods enhanced by high-throughput data collection from patients, can improve disease diagnosis by combining multiple complementary information sources. For example, pathologists use hematoxylin and eosin (H & E) stained tissue images to diagnose diseases or disease subtypes, and H & E image digitization has opened the way to digital pathology. New algorithms have been developed to resolve image features and aid in the diagnosis or study of molecular mechanisms of disease. However, these algorithms do not always capture complex histological heterogeneity. Furthermore, tissue images only give a high level of disease pathology. To address these problems, we propose a method of resolving H & E images to capture interpretable heterogeneity characteristics of tissues such as breast, lung and brain tissues. In some experiments, we applied our analysis pipeline to new data sets we collected from patients with Idiopathic Pulmonary Fibrosis (IPF) and from control subjects.

Formalin-fixed paraffin-embedded (FFPE) lung tissue was collected from university of pittsburgh and Lung Tissue Research Consortium (LTRC) according to a study protocol approved by the institutional review board. A multidisciplinary group of pathologists and clinicians examined the histopathological, clinical and demographic characteristics of subjects in the study and confirmed IPF diagnosis. IPF (n-23) and control (n-7) tissues were used as training groups in this study, while IPF (n-10) and control (n-10) tissues were used as validation groups.

All H & E stained tissues were scanned with an Aperio tissue scanner (lycra biosystems) and submitted for image analysis. The tissue slices in the data set were scanned with a 20 x objective at a resolution of 0.5 pm/pixel, which produced a digital full-slice image (WSI) of approximately 20000 x 30000 pixels. After the tissue object decomposition step, as can be seen in fig. 21A to 21F, each image is represented by approximately 1 million of four different types of tissue objects (lumen, dark stroma, light stroma, and nucleus).

Fig. 21A-21F depict a computational pipeline that models architectural features of histological tissues (e.g., spatial tissues). In particular, fig. 21A to 21B illustrate color normalization. Histopathological data in the form of a full-slice image (WSI) can have a wide range of color appearances due to biological differences, differences in slice preparation techniques, differences in imaging hardware, and the like. We can reduce this variability by preprocessing the original WSI 2102 (whose cropped version is shown in fig. 21A) with the cytological component nuclei 2104 (dark purple regions), the stroma 2106 (regions with pink shading), and the lumen 2108 (white) into the color normalized version 2110 shown in fig. 21B. Fig. 21C to 21D illustrate object map generation: the WSI 2102 is then broken down into circular original objects, i.e., tissue objects 2112, to approximately represent the cytological components of the tissue, as shown in fig. 21C. A neighborhood map is constructed over the tissue object center of the entire WSI by dironey (Delaunay) triangulation 2114, where neighboring objects are connected by edges, as shown in fig. 21D. Fig. 21E to 21F illustrate spatial statistics and feature extraction. We perform random walk on the organizational object graph 2116 shown in fig. 21E to encode the spatial neighborhood statistics around each object. As described further below, a graph, such as graph 2116, is constructed for each of the organizational objects 2112. The tissue objects 2112 are then represented by 100-dimensional vectors representing the spatial characteristics of their neighborhoods, as shown in FIG. 21F. For example, spatial property #31, corresponding to vector element #31, represents the nucleus-to-nucleus interaction at a four-hop distance for the central tissue object.

Analysis of tissue architecture by pathologists is one of the key aspects of many disease diagnosis decisions and is essential for the diagnosis of IPF. Therefore, we developed an image processing and analysis pipeline to capture spatial tissues of different cells within lung tissue. First, we localize cellular components (such as the nucleus) and extracellular components (such as the stroma and air-filled alveolar space). In various embodiments of our technique, instead of segmenting the cytological tissue components, we decompose the purple, pink and white pixels of the image corresponding to the nucleus, stroma and lumen, respectively, into the original circular tissue objects 2112, which are then used as an approximate representation of the cytological components of the tissue. In the present discussion, reference to a "lumen" generally refers to a lumen or white region of tissue.

Since the decomposition is performed on the colors of histopathology images having a wide range of color appearances due to the various factors described above, we use a color normalization method to pre-process the digital tissue images. This is an important step for robust and correct analysis of histopathological images. For our dataset we used the known WSI scalable color normalization method. After color normalization (fig. 21B), in some embodiments, we consider the different shades of pink observed in the matrix region, and we cluster the image pixels into four different groups (purple, light pink, dark pink, and white) using a k-means clustering algorithm for which the clustering centers are calculated from Principal Component Analysis (PCA) on the color distribution. Next, we fit appropriately sized circular objects 2112 to theseAmong the pixels of the cluster (note that noise is taken into account) to represent the nucleus 2112a, light stroma 2112b, dark stroma 2112C, and lumen 2112d objects of the tissue, as shown in fig. 21C. The circular shape is used for analytical convenience and as a representation of the presumed nucleus and its neighborhood. Each image I is composed of a set of tissue objects o (I) ═ o_i]Is shown in which each o_iFrom its central coordinate (x)_i，y_i) Together with its type t_iEpsilon [ cell nucleus, light matrix, dark matrix, lumen]Representation, which is used as a representation of the original nucleus and its neighborhood.

To capture spatial architectural features, we use Diloney triangulation 2114 based on the coordinates (x) of tissue objects_i，y_i)(fig. 21C to 21D) to construct a neighborhood map 2116. Then, we proceed by organizing each of the objects o_iThe respective organizational objects o are defined by being set to the root 2118 of the tree in the corresponding neighborhood map 2116 and by accessing the root's neighboring objects in breadth-first order using the edges of the neighborhood map (also called Diloney map) 2116_iBreadth-first traversal of (c). We set the maximum depth of traversal to h hops from the root tissue object 2118.

Next, for each depth level, we calculate the probability of finding each type of tissue object, as shown in FIG. 21E. Since we have four tissue object types, there are ten different types of edges for each depth level. We choose h 10 as we observed after testing different distances (e.g., h 5, 10, 15, 20, 25) at which random walk converges for h 10. As a result, for a maximum depth of 10 hops, we have a set of 100 probability values that describe the spatial characteristics of the respective tissue objects. For example, in graph 2116, object o is centered from the center at depth level 1_iThe probability of finding another nuclear object is 42.9% (═ 3/7).

Theoretically, these spatial property vectors capture the distribution of various cytological components and their interactions in a defined neighborhood. Therefore, we cluster these vectors representing neighborhood properties into q clusters to find representative architectural features that organize object neighborhoods. To do this, the principal component of the training data is first calculated and q is chosen such that it will cover 95% of the input variance. After cluster center initialization, each tissue object neighborhood is assigned to its closest cluster. And finally, calculating the proportion of q different types of clusters as the architecture feature vector of the image for each image.

The above-described techniques are not limited to lung tissue and are generally applicable to any type of tissue, such as breast tissue, brain tissue, and the like. In some cases, the substrate may not be subdivided into light and dark substrates. Also, instead of only one type of nuclei as in the analysis of lung tissue described above, there may be different types of nuclei, e.g. three types of nuclei as discussed above in connection with breast tissue. Thus, the number of types of objects in the neighborhood graph and, correspondingly, the number of types of edges in the neighborhood graph may vary. Also, the depth h at which the random walk of the neighborhood map will converge may vary with tissue type. Since the size of the vector representing the neighborhood statistics of each object depends on the number of edge types and the depth h in the neighborhood graph, the size of the vector, and thus the number q of clusters of vectors that can approximately (e.g., over 85%, 90%, 95%, etc.) account for input variance, can also vary with tissue type.

Embodiments of HistoMapr may identify Architectural Features (AF) consistent with lung pathophysiology. In our experiments, we captured spatial features from WSI using images as a whole without any cropping or tiling. To do this, we mask the tissue region to avoid white areas around the tissue and construct a neighborhood map, computed spatial properties within the tissue region, and architectural features that summarize the capture of tissue structures as described above. After capturing neighborhood statistics and clustering the decomposed objects into architectural features based on these spatial statistics, we found a total of 12 architectural features covering 95% of the total input variance, i.e., q-12 in this experiment. As noted above, q may vary with tissue type.

Fig. 22A shows a representative example of tissue features. The general pattern depicted is consistent with what we expect from a surgical pathology perspective. Examples are from IPF or normal samples, depending on which is the main category for the particular feature. For the following annotation of AF, the pathologist also examines the surrounding area. In fig. 22B, 12 AFs as color-coded are schematically shown.

In fig. 22A, the example in panel 1 corresponds to AF #1 (mainly in IPF) and shows fibroblast foci. The example in panel 2 corresponds to AF #2, with oligocellular fibrosis appearing to be juxtaposed to normal alveolar lung parenchyma. This may mean various situations: sub-pleural scarring, smoking-related interstitial fibrosis (i.e. scarring in smokers, which is commonly seen in lung backgrounds of cancer resections), IPF regions, regions of fibrotic lung disease of other etiology (e.g. connective tissue disease, allergic pneumonia, etc.), and the like. A related feature is an anucleated fibroblast lesion. This abrupt transition will support the UIP pattern above the NSIP pattern. In control patients without diffuse fibrotic lung disease, this represents nonspecific scarring.

The image example in panel 3 corresponds to AF #3 and depicts a substantially normal lung. Here the lung is collapsed. The red blood cells 2202 are artificial (i.e., due to the process of obtaining the tissue). Our technique may exclude the red region 2202 prior to image feature extraction. The example in panel 4 corresponds to AF #4 and represents adventitial fibrosis around the pulmonary artery branch in the distal bronchial vascular bundle. This phenomenon is common and generally increases with age. The connective tissue surrounding the bronchial vascular bundle is important to allow the lungs to stretch around and highly variable vascular loads under different motor conditions. Note how there is a likewise abrupt transition to the normal surrounding lung, as seen in the example of panel 2 corresponding to AF #2, but this has a medium sized blood vessel in the center, indicating that this is part of a different structure. Blood vessels are not really normal. They are small and somewhat disorganized. The black pigment appears to be carbon dust.

The example of panel 6 corresponds to AF #6 and shows the characteristics of cellular and fibrotic interstitial pneumonia with a nonspecific interstitial pneumonia (NSIP) pattern. Here, the alveolar septum is more or less uniformly expanded by fibroblasts, inflammatory cells, and collagen. There is no aggravation of scarring in the periphery of this lung leaflet (which is delimited by the upper right pleura), but rather scarring involves the leaflet quite diffusely. The air cavity is not significant. In the panel, i think i see some eosinophils, especially in the upper right quadrant. This was a non-specific finding. Tissue eosinophilia can be seen in patients with allergic pneumonia, underlying connective tissue or other systemic diseases, drug toxicity (meaning chronic fibrotic interstitial pneumonia due to side effects of drug treatment), chronic eosinophilic pneumonia (which, if such scarring, would be long-lived). When we see this feature in UIP patients, we propose the possibility that their UIP may be due to connective tissue disease (rather than idiopathic ═ IPF).

An example in the panel 7 corresponding to AF #7 is the characteristics of IPF. This example is completely remodeled in configuration and shows the fibrotic lung parenchyma, which appears to be a honeycomb cyst in the lower portion. Distended air cavities with concentrated mucus are typical of IPF lungs. The amount of chronic inflammation in peripheral fibrosis is also within the normal spectrum of IPF. The epithelium, outside of one large cyst at the base, usually has the eosinophilic appearance of lung cells reactive in pneumonia(s).

The example in panel 8 corresponds to AF #8 and shows respiratory bronchiolitis, characterized by smoking-followed air cavity staining macrophages. The alveolar septa are mostly normal, although some show minimal and inconspicuous thickening. More prominent is the leaflet spacing extending from top to bottom in the left half of the image. This is slightly higher than normal fibrosis and is not very cellular. In addition to macrophages, the air cavities are normal. The most common contents are pigments from smoking, pigments from mineral dust exposure and hemoferritin from blood. Blood in the air cavity occurs in bronchiectasis, infection/ARDS, previous biopsies (this hemosiderosis is common in lung allograft biopsy because one takes multiple biopsies).

The example in panel 11 corresponding to AF #11 depicts a normal lung. The alveolar septa are thin and fragile, and the capillaries do not protrude. There was no alveolar septal edema. The example in panel 12 corresponding to AF #12 similarly depicts a normal lung. The alveolar septum is normal. The air cavity contains some colored macrophages. The arterioles near the upper left of the image edge look normal (no onion skin or muscle hypertrophy).

Thus, fig. 22A illustrates that the above-described architectural features can distinguish between normal and fibrotic lung regions. In particular, panels 1-12 of fig. 22A show that architectural features derived from the neighborhood of cytological tissue components may help to differentiate IPFs. These panels show prototype regions of various architectural features. For example, panel 1 shows the main area of nuclear to light stroma interaction, represented by AF # 1; panel 11 shows the major areas of cell nucleus to lumen and lumen to lumen interaction; these regions are commonly annotated by a pathologist as fibroblast foci and normal alveolar regions, respectively.

Fig. 22B shows the relative population of architectural features separated for IPF and control groups. Note that architectural features #1 and #7 are dominant in the IPF image, while features #8 and #11 are dominant in the control image. To visualize some of these patterns in the WSI to be analyzed, an embodiment of HistoMapr highlights them by the color assigned to the individual AFs. In fig. 22B, the assigned colors are shown below the x-axis. The objects in the WSI are colored according to the features as shown in fig. 22C to 22D. In particular, fig. 22C shows an example WSI from an IPF patient, with features # 12212 (blue) and # 72214 (dark green) predominating. Fig. 22D shows an example WSI from the control group, where features # 22216 (light green) and # 112218 (pink) are dominant.

The frequency percentages of these 12 features are shown in fig. 22B, where the difference between IPF and control types is clearly visible, especially for AF #1, #7, #8 and #11 (i.e. features #1 and #7 are predominant in fibrotic tissue, while features #8 and #11 are predominant in normal tissue). Note that these architectural features are connected to the organizational structure, as they rely on organizational object interactions. For example, for connections in the random walk graph, AF #1 represents a neighborhood where the nucleus of the tissue object and the light stroma interact significantly more than other types of interactions, and this is followed by interactions between light stroma objects; AF #11 represents a neighborhood where the nucleus and the lumen of a tissue object interact significantly more than other types of interactions, followed by interaction between the luminal objects. Expressing these neighborhood characteristics as architectural features, embodiments of HistoMapr-Lung can generalize individual WSIs based on how often these features are detected. By comparing fig. 22C and 22D, differences in characteristics between the IPF and the control sample can be easily observed.

HistoMapr-Brain: analysis of brain tissue

Some embodiments of HistoMapr, referred to as HistoMapr-Brain, can analyze Brain tissue. In our experiments, we collected 99 WSIs of brain tissue from Ivy GAP, which contains a Cellular Tumor (CT) region. Here, the obtained WSI has been H & E stained normalized for color. However, if an unprocessed WSI is received, then staining and color normalization will be performed in order to robustly and correctly analyze the WSI. The tissue section from which the WSI data set was generated was scanned with a 20 x objective at a resolution of 0.5 pm/pixel, which resulted in a digital WSI of approximately 15000 x 18000 pixels. The received images are also semi-automatically annotated by the Ivy GAP to identify certain anatomical features in the images. Annotations included Leading Edge (LE), Invasive Tumor (IT), Cellular Tumor (CT), and Necrosis (NE). For our experiments, we extracted the architectural features of the image from the CT region of WSI via the above process. Using the above process, architectural features of other areas of the WSI can also be derived. Two examples of WSIs with CT regions annotated with green are depicted in fig. 23A to 23D.

In particular, representative glioblastoma tissue sections of the conventional and proto-neuron types 2302C, 2302p are shown in fig. 23A and 23C, respectively. In referring to fig. 23A to 23D, the letter "c" is associated with a conventional image/region/feature, and the letter "p" is associated with a primary nerve image/region/feature. Fig. 23B and 23D show the annotation region of each of these images as: cellular Tumor (CT) regions 2304c, 2304p (green); infiltrative Tumor (IT) regions 2306c, 2306p (magenta); leading Edge (LE) regions 2308c, 2308p (blue); necrotic (N) regions 2310c, 2310p (black); and a proliferating blood vessel (CThbv) region 2312p (orange) in cellular tumors.

The HistoMapr-Brain embodiment decomposes the object graph based on CT region masking using annotations provided by Ivy GAP. We then construct a neighborhood map and capture the spatial statistics within these masked CT regions, as described above in connection with the overall HistoMapr computation pipeline. After capturing neighborhood statistics and clustering the decomposed objects into architectural features based on these spatial statistics, we find a total of 14 architectural features that cover 95% of the total input variance (i.e., q-14).

Fig. 24A shows the frequency percentage of these 14 features, where the difference between the afferent and the proneural nerve types is clearly visible, especially for features #1, #2 and # 6. Features #1 and #6 were predominant in the primary neural type cell tumor region, while feature #2 was predominant in the classical type cell tumor region.

Specifically, fig. 24A is a histogram of architectural features, which are numbered 1 through 14 and color coded. The frequency of representative classical and protoglioblastoma subtypes is different for some characteristics. For example, note the relatively high population of features #1 and #6 in the proto-neural subtype and the relatively high population of feature #2 in the classical subtype. Then, as shown in fig. 24B, an embodiment of HistoMapr-Brain is identified in the respective CT regions 2304c, 2304p of WSI feature #2 by marking the corresponding object 2402c in red (which is the assigned color of feature # 2). Likewise, as shown in fig. 24C, an embodiment of HistoMapr-Brain is also identified in the respective CT regions 2304C, 2304p of WSI features #1 and #6 by marking the corresponding objects 2404p, 2406p green and dark blue, respectively, which are the respective assigned colors for features #1 and # 6. By comparing fig. 24B and fig. 22C, the difference in labeling characteristics between the conventional WSI region and the protogenic WSI region can be easily observed. In traditional WSI, architectural features 2402 (red) are dominant, while in the proto-neural WSU, architectural features 2404p (green) and 2406p (dark blue) are dominant.

Appendix A

Determination of the region of interest:

we characterize the spatial arrangement of nuclei in tissue images and use this information to segment regions of interest (ROI) (see fig. 25A to 25F). The various steps involved in segmenting the ROI include:

1. in WSI 2500 (fig. 25A, top left), we identified putative nuclear locations in the form of superpixels (e.g., circular objects) in hematoxylin-stained tissue regions (H), and used superpixels derived from eosin-stained (E) and white tissue regions (W) to characterize neighborhood statistics. In some embodiments, after color normalization, the H super pixel 2502 is labeled purple, the E super pixel 2504 is labeled pink, and the W super pixel 2506 is labeled cyan. (FIG. 25B, upper right, and FIG. 25C, upper right).

2. To characterize the nuclear neighborhood, we combine superpixels from all three classes H, E and W and perform dironey triangulation using their center coordinates (fig. 25D, bottom left).

3. We calculate the pair-wise distance between superpixels of the same class. Then, for each class, we construct a separate graph in which each superpixel as a node and the neighboring superpixels of that class are connected by an edge when their distance is below a specified threshold τ.

4. We set the distance threshold τ to be at least the median of the distance distribution between neighboring superpixels (fig. 25E, bottom middle).

5. We use a greedy connected component analysis algorithm to cluster superpixels into fragments.

6. We sort the fragments in the H and W categories in descending order based on their tissue regions. We combine fragments from both classes with a simple post-processing rule: if the segment W_iAnd fragment H_jOverlap, the overlap portion is absorbed into H regardless of the overlap region_jIn (1).

7. After merging the fragments from the H and W classes, we get an optimized segmentation of the ROIs 2508, 2510 (fig. 25F, bottom right).

Ranking of regions of interest:

ROIs may be ranked based on their cytological and architectural featuresName (name)The following are described:

1. to quantify the cytological characteristics, we first segment the nuclei by generating a more accurate set of nuclear masks in the respective ROIs. Segmentation in this context typically involves applying a threshold to the hematoxylin color channel and obtaining a putative nuclear region. Watershed can be used to separate touching and overlapping nuclei, and morphological operations can be used to fill any remaining holes. Next, small and large segmented objects and objects near the image boundaries can be eliminated. Finally, a round of erosion is performed followed by dilation to smooth the nuclear shape. The cell nucleus mask is a connected component that represents the pixels of the individual cell nuclei individually and more accurately. These masks help to locate the shape characteristics and center coordinates of the nucleus in question.

2. After nuclear segmentation, we calculated 196 nuclear features for each nucleus in the ROI.

3. For breast lesions, we found three dominant cytological phenotypes, which we captured using the k-means clustering algorithm. k-means clustering was used to identify dominant cytological phenotypes. Once they are identified, xAI will simply use them for some analysis. If different types of analysis are to be performed, for example diagnosis of skin cancer, different cytological phenotypes may be correlated/dominant and need to be identified, which can be done using k-means clustering. Thus, in general, different numbers and/or different types of dominant cytological phenotypes may exist for different tissue conditions. These steps may be repeated for different numbers and/or types of dominant cytological phenotypes.

4. These phenotypes may be the result of normal, atypical and polymorphic nuclei in high-risk benign breast lesions (fig. 2C). The amount and type of these cytological phenotypes may vary for different types of tissues.

5. For the task of ROI ranking, we constructed a Cytological Feature (CF) vector for each ROI that was the aggregate statistics (e.g., mean, median, standard deviation, etc.) of each of the 196 features across all nuclei in the ROI under consideration.

6. To quantify the architectural properties, we capture the spatial properties of the organizational content.

7. Principally, the ROI is represented by 5 different objects: nuclei (nucleoli) from three cytological phenotyping of previous steps₁、nuclei₂、nuclei₃) And two superpixel-based components (stroma and lumen) from the ROI segmentation step, as shown in fig. 2C. Here, nucleoli _ l, nucleoli _2 and nucleoli _3 are computational phenotypes, which may be assigned histological meaning by a professional pathologist. Typically, the lumen superpixels correspond to the voids and openings, and the stromal superpixels correspond to the connected tissue. These superpixels do not overlap with the nuclei.

8. In this dataset, polymorphic phenotypes are shown as distinct subgroups.

9. To characterize the neighborhood around each object, a spatial network is constructed by breadth-first traversal for a small number of depth levels from each object (fig. 2D).

10. At each depth level, we compute the finding 15 (this is influenced by the number of object types defined in the previous step, which would be 6 for 3 types of objects) different object connections (e.g. cell nuclei)₁(nuclei₁) -cell nucleus₁(nuclei₁) Cell nucleus₁(nuclei₁) Stroma, nucleus₁(nuclei₁) Lumen, etc.). In some embodiments, an object connection is an edge in a dironi triangulation connecting two objects that are neighbors. As a result, for a maximum depth of 5, we generate a set of 75 probability values describing the neighborhood statistics of the respective object, i.e. the probability that another nucleus of type nucleous _1, nucleous _2 or nucleous _3 is close/adjacent to nucleus-a, or stroma is close/adjacent to nucleus-a, or lumen is close/adjacent to nucleus-a, wherein the probability is calculated separately for the respective depth level. Each depth level represents a set of neighbors at a given distance, where the distance passes from on the graphThe number of hops or jumps taken by the central nucleus. For each hop, another depth level may be reached in a breadth first search manner.

11. For phenotypical spatial networks, we clustered neighborhood statistics into q clusters by noting the main subspace that captured 95% of the input variance. Here, the CF vector defines the nuclear phenotype. Any suitable range of candidate values for q may be considered. In one study, q varied from 5 to 38.

12. The architecture phenotype is learned by applying a k-NN (k nearest neighbors) algorithm, where for each candidate q, k equals q.

13. The respective ROIs are then represented by the relative proportion of the q architectural patterns.

In some embodiments, we perform random walk with the architecture pattern as a new object, and perform the new object in a layered fashion, with the goal of spatially characterizing a full-slice tissue image.

14. We construct three additional sets of architectural feature vectors based on:

a. color-based architectural features (AF-C) using superpixel derived nuclei, stroma, and luminal objects;

b. architectural features based on cytological phenotype nuclei (AF-N), which use nuclear phenotype alone; and

c. combined architectural features (AF-CN) using nuclear phenotypes in combination with matrix and luminal superpixels.

15. In our preliminary study data (46 breast lesion cases, 93 full-slice images), we found q ═ 18 patterns/features, including architectural and/or cytological patterns/features representing ROIs (fig. 3).

16. We used logistic regression to classify/rank the ROIs based on the feature sets CF, AF-C, AF-N, and AF-CN.

Shunting:

individual ROI markers have diagnosis and confidence scores.

The ROIs are typically sorted in descending order based on 1) the severity of the diagnostic label and/or 2) the confidence score, such that the most confident, most severe labeled ROI is typically ranked above the less confident ROI or the less severe ROI.

Shunting after each ROI has a label or rank.

Temporary diagnosis

HistoMapr is not a diagnostic system. It is an ancillary system designed to guide pathologists during their daily effort to log out cases.

HistoMapr previews cases and then dynamically assigns cases to one or more pathologists based on case difficulty (expert specialist versus general pathologist) or workload (even distribution of pathological cases among pathologists).

2. The pathologist logs in to access the cases assigned to them, which appear as a table with orderable fields including case ID, HistoMapr status, and additional information about the case (e.g., organs, date of treatment, markers about urgency, revisiting patients, etc.).

3. They opened a case that would bring the HistoMapr case view interface (fig. 4). The xAI platform may rank the fields/ROIs according to urgency (e.g., more difficult or ambiguous cases, simple, high confidence cases, etc.).

4. In the middle of the interface is a full-slice image viewer-the pathologist always controls the case.

5. Standard and familiar controls of the basic user; the breast core biopsy image of the case under examination is in the right panel and there is patient information in the left panel.

HistoMapr previews all WSIs and finds diagnostically relevant ROIs, and modifies the border pixels to mark the ROIs, at the bottom panel, ROIs collected from all WSIs of the case are shunted from malignant to benign (left to right).

7. For each ROI, the HistoMapr selected provided diagnostic labels with key differences: the releasable AI-HistoMapr has "why? "button that provides a diagnostic explanation for each ROI with a label explaining why HistoMapr makes a decision about that label and the associated confidence score.

HistoMapr by its marking of ROI, with "why? "information supports these tags to provide decision support; this reduces uncertainty and improves diagnostic confidence.

9. The pathologist has several workflow options-the consent/non-consent box means work is supported for the pathologist. When the pathologist wishes to end the HistoMapr review, a "done" button is used to present a summary of the assessment.

10. The summary page is to show the pathologist an audit of the ROIs that have been observed and the labels that have been approved/disapproved/possible (fig. 5). The disapproved/possible ROIs can be further reviewed and new labels given using the HistoMapr annotation tool described in the original document "innovative features, item 2".

In response to a user activation "why? "interface provides explanation

The working order of the generative reasoning/interpretation of the ROI label is:

1. the ROI is segmented.

2. Descriptive spatial features are computed for each ROI.

3. The features are further phenotyped based on their similarity to a main subspace describing the architectural pattern.

4. The architecture patterns from the other images were visually examined by an expert (pathologist) and assigned histopathologically meaningful descriptions (i.e., stiffness, matrix density, proliferation, etc.) before analyzing new, previously unanalyzed WSIs using the xAI platform.

5. Classification/labeling of the ROI based on relative architecture pattern ratios of the ROI using logistic regression (classification can be done using other methods such as SVM, random forest, etc.).

6. The ROIs are sorted based on their severity level (the severity of each label is pre-defined in the literature by the pathologist). The sorting can also be performed in two steps: either step may be performed before the other based on the severity level in step a and the confidence score in step B.

7. The distribution of architectural and/or cytological patterns/features within individual ROIs is quantified and their presence/absence (i.e., low/medium/high stiffness, etc.) is reported in the interpretation interface. Each pattern feature may have a positive/negative indicator indicating how the pattern/feature contributes to the ROI diagnostic label. This provides the basis for the confidence score.

It should be clear that there are many ways of constructing the apparatus and/or system components, interfaces, communication links, and methods described herein. The disclosed methods, apparatus, and systems may be deployed on convenient processor platforms, including web servers, personal and portable computers, and/or other processing platforms. Other platforms may be expected to improve processing power, including personal digital assistants, computerized watches, cellular telephones, and/or other portable devices. The disclosed methods and systems may be integrated with known network management systems and methods. The disclosed methods and systems may operate as SNMP agents and may be constructed with the IP address of the remote machine running the appropriate management platform. Accordingly, the scope of the disclosed methods and systems is not limited by the examples given herein, but may include the full scope of the claims and their legal equivalents.

The methods, apparatus and systems described herein are not limited to a particular hardware or software configuration and may find application in many computing or processing environments. The methods, apparatus and systems may be implemented in hardware or software or a combination of hardware and software. The methods, apparatus and systems may be implemented in one or more computer programs, where the computer program(s) may be understood to include one or more processor-executable instructions. A computer program may be executed on one or more programmable processing elements or machines and may be stored on one or more storage media (including volatile and non-volatile memory and/or storage elements), one or more input devices, and/or one or more output devices that are readable by a processor. Thus, a processing element/machine may access one or more input devices to obtain input data and may access one or more output devices to communicate output data. The input and/or output means may comprise one or more of: random Access Memory (RAM), Redundant Array of Independent Disks (RAID), floppy disk drive, CD, DVD, magnetic disk, internal hard disk, external hard disk, memory stick, or other storage device accessible by a processing element as provided herein, wherein such aforementioned examples are not exhaustive and are for purposes of illustration and not limitation.

The computer program(s) can be implemented using one or more high-level procedural or object-oriented programming languages to communicate with a computer system; however, the program(s) can be implemented in assembly or machine language, if desired. The language may be compiled or interpreted. The sets and subsets typically include one or more members.

Thus, as provided herein, the processor(s) and/or processing elements may be embedded in one or more devices operating independently or together in a networked environment, where the network may include, for example, a Local Area Network (LAN), a Wide Area Network (WAN), and/or may include an intranet and/or the internet and/or another network. The network(s) may be wired or wireless or a combination thereof and may use one or more communication protocols to facilitate communication between the different processors/processing elements. The processor may be constructed as a distributed process, and in some embodiments may use a client-server model as needed. Thus, methods, apparatus and systems may use multiple processors and/or processor devices, and processor/processing element instructions may be divided among such single or multiple processors/devices/processing elements.

The apparatus (es) or computer system(s) integrated with the processor (s)/processing element(s) may include, for example, a personal computer(s), a workstation (e.g., dell, hewlett packard), a Personal Digital Assistant (PDA), a handheld device such as a cellular telephone, a laptop computer, a palmtop computer, or another apparatus capable of being integrated with the processor(s) that may operate as provided herein. Accordingly, the devices provided herein are not exhaustive and are provided for illustration and not limitation.

References to "processor" or "processing element," "the processor," and "the processing element" may be understood to include one or more microprocessors that may communicate in stand-alone and/or distributed environment(s) and thus may be configured to communicate with other processors via wired or wireless communication, where such one or more processors may be configured to operate on a device that may be controlled by one or more processors/processing elements of similar or different devices. Thus, use of such "microprocessor," "processor," or "processing element" terms may also be understood to include a central processing unit, arithmetic logic unit, application specific Integrated Circuit (IC), and/or task engine, such examples provided for illustration and not limitation.

Further, unless otherwise specified, references to memory may include one or more processor-readable and accessible memory elements and/or components that may be internal to the processor control device, external to the processor control device, and/or accessible via a wired or wireless network using various communication protocols, and unless otherwise specified, may be provided to include a combination of external and internal memory devices, where such memory may be contiguous and/or application partition based. For example, the memory may be a flash drive, computer disk, CD/DVD, distributed memory, or the like. References to structures include links, queues, graphs, trees, and such structures are provided for illustration and not limitation. References herein to instructions or executable instructions may be understood to include programmable hardware in light of the above.

Although the methods and systems have been described with respect to specific embodiments thereof, they are not so limited. It will thus be seen that numerous modifications and variations are possible in light of the above teachings. Many additional variations in the details, materials, and arrangements of parts described and illustrated herein may be made by those skilled in the art. Thus, it will be understood that the methods, apparatus and systems provided herein are not limited to the embodiments disclosed herein, may include practices other than those specifically described, and are to be interpreted in the broadest scope permitted by law.

Claims (62)

1. A method for performing interpretable pathology analysis of a medical image, the method comprising:

for a region of interest (ROI) in a full-slice image (WSI) of tissue, identifying features of a plurality of feature types, wherein at least one feature type is at least partially indicative of a pathological condition of the tissue within the ROI;

using a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of the plurality of feature types: (i) classifying the ROI into a category within the plurality of categories; and (ii) assigning a label to the ROI, the label indicating a tissue condition associated with the category and the tissue in the ROI;

storing the specified explanatory information about the tag, the explanatory information including information about the identified features; and

displaying: (i) at least a portion of the WSI in which a boundary of the ROI is highlighted; (ii) the label assigned to the ROI; and (iii) a User Interface (UI) comprising: (a) a first UI element for providing a user with access to the stored explanatory information; and (b) one or more additional UI elements that enable the user to provide feedback on the specified tags.

2. The method of claim 1, wherein,

the tissue comprises breast tissue; and is

The plurality of categories of tissue conditions include two or more of: invasive cancer, Ductal Carcinoma In Situ (DCIS), high risk benign, low risk benign, ductal epithelial atypical hyperplasia (ADH), Flattened Epithelial Atypical (FEA), Columnar Cell Change (CCC), and normal duct.

3. The method of claim 1, wherein,

the tissue comprises lung tissue; and is

The plurality of categories of tissue conditions include: idiopathic Pulmonary Fibrosis (IPF) and normal.

4. The method of claim 1, wherein,

the tissue comprises brain tissue; and is

The plurality of categories of tissue conditions include: classical cell tumors and protoneurocyte tumors.

5. The method of claim 1, wherein a feature type is a cytological feature or an Architectural Feature (AF).

6. The method of claim 5, wherein the characteristic type cytological characteristic is characterized by one of the following subtypes: nuclear size, nuclear shape, nuclear morphology, or nuclear texture.

7. The method of claim 5, wherein the feature of the feature type architecture feature is one of the following subtypes: an architectural feature (AF-C) based on colors of a set of superpixels in the ROI; (ii) architectural features (AF-N) based on a cytological phenotype of nuclei in the ROI; or (iii) a combined architectural feature (AF-CN) based on both the color of a set of superpixels in the ROI and the cytological phenotype of nuclei in the ROI.

8. The method of claim 5, wherein the feature of the feature type architecture feature is one of the following subtypes: nuclear arrangement, stromal cellularity, epithelial pattern in ducts, epithelial pattern in glands, cell cobblestone, stromal density, or hyperplastic.

9. The method of claim 1, wherein the information about the features comprises one or more of:

a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label;

a count of features of a particular feature type detected in the ROI;

a measured density of features of the particular feature type in the ROI; or

A strength of the particular feature type in indicating the condition of the tissue.

10. The method of claim 1, wherein the explanatory information includes a confidence score calculated by the classifier in specifying the label, wherein the confidence score is based on one or more of:

a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label;

for a first feature type: (i) a strength of the first feature type in indicating the condition of the tissue; or (ii) a count of features of the first feature type detected in the ROI; or

Another total number of feature types detected in the ROI but corresponding to tissue conditions other than the condition associated with the label.

11. The method of claim 1, further comprising:

in response to the user interacting with the first UI element:

generating an explanatory description using a standard pathology vocabulary and the stored explanatory information; and is

The explanatory description is displayed in an overlay window, side panel, or page.

12. The method of claim 11, further comprising:

highlighting features of a particular feature type in the ROI using a color assigned to the particular feature type, the particular feature type being at least partially indicative of the tissue condition indicated by the label; and

displaying the highlighted ROI in the overlay window, the side panel, or the page.

13. The method of claim 1, further comprising:

repeating the identifying, designating and storing steps for a plurality of different ROIs; and

prior to the displaying step, (i) calculating a respective risk metric for each of the ROIs, wherein the risk metric for a ROI is based on: (a) a designation label of the ROI; or (b) a confidence score for the ROI; and (ii) ordering the ROIs according to the respective risk metrics of the ROIs,

wherein the displaying step comprises:

displaying in one panel: (i) at least a portion of the WSI with the boundary of the ROI having the highest risk metric highlighted; (ii) the label assigned to the ROI; and (iii) a User Interface (UI) providing the user with access to the stored interpretation of the specified label for the ROI; and

thumbnails of the ROI sequence are displayed in another panel.

14. The method of claim 1, further comprising:

obtaining the full-slice image (WSI); and

identifying the ROI in the WSI, wherein identifying the ROI comprises: (i) marking at least two types of superpixels in the WSI, one type corresponding to hematoxylin stained tissue and the other type corresponding to eosin stained tissue; and (ii) marking segments of pixels of the first type to define an occlusion region as the ROI.

15. The method of claim 14, further comprising: identifying a plurality of ROIs in the WSI.

16. The method of claim 1, further comprising: updating a training data set for the classifier, the updating the training data set comprising:

receiving feedback from the user via the one or more additional UI elements for the label designated for the ROI, wherein the feedback indicates a correctness of the designated label; and

storing a portion of the WSI associated with the ROI and the specified label in a training dataset.

17. The method of claim 1, wherein the classifier is selected from the group consisting of: decision trees, random forests, support vector machines, artificial neural networks, and classifiers based on logistic regression.

18. A method for distributing cases among a group of pathologists, the method comprising:

for each of a plurality of cases, processing a corresponding full-slice image (WSI) of tissue, the processing of the WSI comprising:

identifying one or more regions of interest (ROIs) in the WSI, each ROI assigned a respective diagnostic label indicative of a condition of tissue in the ROI;

for each ROI, calculating a respective confidence score for the respective designation;

calculating, for the WSI: (i) a severity score based on the respective diagnostic label assigned to the one or more ROIs in the WSI; and (ii) a confidence level based on the respective confidence scores of the one or more ROIs;

storing the severity score, the confidence level, and the respective confidence score as explanatory information;

sending the WSI to an emergency pathologist in the group of pathologists if the severity score is equal to or above a specified threshold severity score;

otherwise, if the confidence level is equal to or below a specified threshold confidence level, sending the WSI to a second specialist in the group of pathologists; and

otherwise, the case is sent to a general pathologist in the family of pathologies.

19. The method of claim 18, wherein sending the case to a general pathologist comprises:

selecting an ordinary pathologist from a library of ordinary pathologists within the set of pathologists such that a balanced workload of the library is maintained when sending the case to the selected ordinary pathologist.

20. The method of claim 18, further comprising: assigning the respective diagnostic label to at least one ROI in at least one WSI, the assigning the respective diagnostic label to a particular ROI comprising:

using a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of a plurality of feature types identified in the image:

classifying the ROI into a category within a plurality of categories; and is

Assigning a label to the ROI, the label indicating a tissue condition associated with the category.

21. The method of claim 18, wherein,

for at least one ROI in at least one WSI, the corresponding diagnostic label is provided by a previous reviewer; and is

The group of pathologists represents a subsequent group of reviewers.

22. The method of claim 18, further comprising:

in response to a user requesting an interpretation of a particular WSI's transmission via a UI element:

generating an explanatory description using the standard pathology vocabulary and the stored explanatory information; and is

Displaying the explanatory description.

23. The method of claim 22, further comprising:

selecting an ROI from the particular WSI for which the specified label indicates a severe condition or for which the confidence score is equal to or below a specified threshold confidence score;

highlighting in the ROI the feature of a particular feature type using a color assigned to the particular feature type, the particular feature type being indicative, at least in part, of the tissue condition indicated by the label assigned to the ROI; and

displaying the highlighted ROI along with the explanatory description.

24. The method of claim 18, wherein, for a first ROI in a first WSI, the explanatory information includes one or more of:

a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI;

a count of features of a particular feature type detected in the first ROI;

a measured density of features of the particular feature type in the first ROI; or

A strength of the particular feature type in indicating a corresponding tissue condition.

25. The method of claim 18, wherein the confidence score for a first ROI in a first WSI is based on one or more of:

a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI;

for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of the first feature type detected in the ROI; or

Another total number of feature types detected in the first ROI but corresponding to tissue conditions other than conditions associated with the label assigned to the first ROI.

26. A method for truth labeling images used to train a classifier, the method comprising:

obtaining a full-section image (WSI) of the tissue;

identifying one or more regions of interest (ROIs) in the WSI, wherein identifying an ROI comprises: (i) marking at least two types of superpixels in the WSI, one type corresponding to hematoxylin stained tissue and the other type corresponding to eosin stained tissue; and (ii) marking segments of pixels of a first type to define an enclosed region as the ROI;

displaying one or more ROIs in sequence; and

for each ROI:

displaying one or more UI elements, wherein a first UI element provides or confirms a respective truth label to be assigned to the ROI; and

in response to a user interaction using the first UI element, the respective truth label is specified to the ROI and the ROI is stored in a training corpus.

27. The method of claim 26, wherein,

the first UI element indicates consent of the user with the provided suggestion, and the method further comprises: for each ROI in at least a subset of the one or more ROIs:

identifying features of a plurality of feature types, wherein at least one feature type is indicative, at least in part, of a pathological condition of the tissue within the ROI;

using a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of a plurality of feature types: (i) classifying the ROI into a category within a plurality of categories; (ii) assigning a suggestion tag to the ROI, the suggestion tag indicating a tissue condition associated with the category; and (iii) storing explanatory information regarding the designation of the suggested tag, the explanatory information including information regarding the identified feature; and

displaying the suggestion tag as the provided suggestion.

28. The method of claim 27, further comprising:

in response to a user requesting interpretation of the suggested tags for a particular ROI via a UI element:

generating an explanatory description using the standard pathology vocabulary and the stored explanatory information; and is

Displaying the explanatory description.

29. The method of claim 28, further comprising:

highlighting features of a particular feature type in the particular ROI using a color assigned to the particular feature type, the particular feature type being indicative, at least in part, of the tissue condition indicated by the label assigned to the ROI; and

displaying the highlighted ROI along with the explanatory description.

30. The method of claim 28, wherein, for the particular ROI, the explanatory information includes one or more of:

a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI;

a count of features of a particular feature type detected in the particular ROI;

a measured density of features of the particular feature type in the particular ROI; or

A strength of the particular feature type in indicating a corresponding tissue condition.

31. The method of claim 28, wherein the ROI-specific confidence score is based on one or more of:

a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI;

for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of the first feature type detected in the particular ROI; or

Another total number of feature types detected in the particular ROI but corresponding to tissue conditions other than conditions associated with the suggestion tag assigned to the particular ROI.

32. A system for performing interpretable pathology analysis of a medical image, the system comprising:

a first processor; and

a first memory in electrical communication with the first processor and comprising instructions that, when executed by a processing unit comprising the first or second processor and in electrical communication with a storage module comprising the first or second memory, program the processing unit to:

for a region of interest (ROI) in a full-slice image (WSI) of tissue, identifying features of a plurality of feature types, wherein at least one feature type is at least partially indicative of a pathological condition of the tissue within the ROI;

operating as a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of the plurality of feature types to: (i) classifying the ROI into a category within a plurality of categories; and (ii) assigning a label to the ROI, the label indicating a tissue condition associated with the category and the tissue in the ROI;

storing the specified explanatory information about the tag, the explanatory information including information about the identified features; and is

Displaying: (i) at least a portion of the boundary of the ROI of the WSI is highlighted; (ii) the label assigned to the ROI; and (iii) a User Interface (UI) comprising: (a) a first UI element for providing a user with access to the stored explanatory information; and (b) one or more additional UI elements that enable the user to provide feedback on specified tags.

33. The system of claim 32, wherein,

the tissue comprises breast tissue; and is

The plurality of categories of tissue conditions include two or more of: invasive cancer, Ductal Carcinoma In Situ (DCIS), high risk benign, low risk benign, ductal epithelial atypical hyperplasia (ADH), Flattened Epithelial Atypical (FEA), Columnar Cell Change (CCC), and normal duct.

34. The system of claim 32, wherein,

the tissue comprises lung tissue; and is

The plurality of categories of tissue conditions include: idiopathic Pulmonary Fibrosis (IPF) and normal.

35. The system of claim 32, wherein,

the tissue comprises brain tissue; and is

The plurality of categories of tissue conditions include: classical cell tumors and protoneurocyte tumors.

36. The system of claim 32, wherein a feature type is a cytological feature or an Architectural Feature (AF).

37. The system of claim 36, wherein the characteristic type cytological characteristic is characterized by being one of the following subtypes: nuclear size, nuclear shape, nuclear morphology, or nuclear texture.

38. The system of claim 36, wherein the feature of the feature type architecture feature has one of the following subtypes: an architectural feature (AF-C) based on colors of a set of superpixels in the ROI; (ii) architectural features (AF-N) based on a cytological phenotype of nuclei in the ROI; or (iii) a combined architectural feature (AF-CN) based on both the color of a set of superpixels in the ROI and the cytological phenotype of nuclei in the ROI.

39. The system of claim 36, wherein the feature of the feature type architecture feature has one of the following subtypes: nuclear arrangement, stromal cellularity, epithelial pattern in ducts, epithelial pattern in glands, cell cobblestone, stromal density, or hyperplastic.

40. The system of claim 32, wherein the information about the features includes one or more of:

a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label;

a count of features of a particular feature type detected in the ROI;

a measured density of features of the particular feature type in the ROI; or

A strength of the particular feature type in indicating the condition of the tissue.

41. The system of claim 32, wherein the explanatory information includes a confidence score calculated by the classifier in specifying the label, wherein the confidence score may be based on one or more of:

a total number of feature types detected in the ROI and corresponding to the tissue condition indicated by the label;

for a first feature type: (i) a strength of the first feature type in indicating the condition of the tissue; or (ii) a count of features of the first feature type detected in the ROI; or

Another total number of feature types detected in the ROI but corresponding to a tissue condition different from the condition associated with the label.

42. The system of claim 32, wherein the instructions further program the processing unit to:

in response to the user interacting with the first UI element:

generating an explanatory description using the standard pathology vocabulary and the stored explanatory information; and is

The explanatory description is displayed in an overlay window, side panel, or page.

43. The system of claim 42, wherein the instructions further program the processing unit to:

highlighting features of a particular feature type in the ROI using a color assigned to the particular feature type, the particular feature type being at least partially indicative of the tissue condition indicated by the label; and is

Displaying the highlighted ROI in the overlay window, the side panel, or the page.

44. The system of claim 32, wherein,

the instructions further program the processing unit to:

repeating the identifying, designating, and storing operations for a plurality of different ROIs; and is

Prior to the display operation, (i) calculating a respective risk metric for each of the ROIs, the risk metric for a ROI based on: (a) a designation label of the ROI; or (b) a confidence score for the ROI; and (ii) ordering the ROIs according to the respective risk metrics of the ROIs; and is

To perform the display operation, the instructions program the processing unit to:

displaying in one panel: (i) at least a portion of the WSI with the boundary of the ROI having the highest risk metric highlighted; (ii) the label assigned to the ROI; and (iii) a User Interface (UI) providing the user with access to the stored interpretation of the specified label for the ROI; and

thumbnails of the ROI sequence are displayed in another panel.

45. The system of claim 32, wherein the instructions further program the processing unit to:

obtaining the full-slice image (WSI); and is

Identifying the ROI in the WSI, wherein, to identify the ROI, the instructions program the processing unit to: (i) marking at least two types of superpixels in the WSI, one type corresponding to hematoxylin stained tissue and the other type corresponding to eosin stained tissue; and (ii) marking segments of pixels of the first type to define an occlusion region as the ROI.

46. The system of claim 45, wherein the instructions further program the processing unit to:

identifying a plurality of ROIs in the WSI.

47. The system of claim 32, wherein the instructions further program the processing unit to:

updating a training data set of the classifier, wherein to update the training data set, the instructions program the processing unit to:

receiving feedback from the user via the one or more additional UI elements for the label designated for the ROI, wherein the feedback indicates a correctness of the designated label; and is

Storing a portion of the WSI associated with the ROI and the specified label in a training dataset.

48. The system of claim 32, wherein the classifier is selected from the group consisting of: decision trees, random forests, support vector machines, artificial neural networks, and classifiers based on logistic regression.

49. A system for distributing cases among a group of pathologists, the system comprising:

a first processor; and

a first memory in electrical communication with the first processor and comprising instructions that, when executed by a processing unit comprising the first or second processor and in electrical communication with a storage module comprising the first or second memory, program the processing unit to:

for each of a plurality of cases, processing a corresponding full-slice image (WSI) of tissue, wherein to process the WSI, the instructions program the processing unit to:

identifying one or more regions of interest (ROIs) in the WSI, each ROI assigned a respective diagnostic label indicative of a condition of tissue in the ROI;

for each ROI, calculating a respective confidence score for the respective designation;

calculating, for the WSI: (i) a severity score based on the respective diagnostic label assigned to the one or more ROIs in the WSI; and (ii) a confidence level based on the respective confidence scores of the one or more ROIs;

storing the severity score, the confidence level, and the respective confidence score as explanatory information;

sending the WSI to an emergency pathologist in the group of pathologists if the severity score is equal to or above a specified threshold severity score;

otherwise, if the confidence level is equal to or below a specified threshold confidence level, sending the WSI to a second specialist in the group of pathologists;

otherwise, the case is sent to a general pathologist in the family of pathologies.

50. The system of claim 49, wherein to send the case to a general pathologist, the instructions program the processing unit to:

selecting an ordinary pathologist from a library of ordinary pathologists within the set of pathologists such that a balanced workload of the library is maintained when sending the case to the selected ordinary pathologist.

51. The system of claim 49, wherein the instructions further program the processing unit to:

assigning the respective diagnostic label to at least one ROI in at least one WSI, wherein to assign the respective diagnostic label to a particular ROI, the instructions program the processing unit to:

operating as a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of a plurality of feature types identified in the image to:

(i) classifying the ROI into a category within a plurality of categories; and (ii) assigning a label to the ROI, the label indicating a tissue condition associated with the category.

52. The system of claim 49, wherein,

for at least one ROI in at least one WSI, the corresponding diagnostic label is provided by a previous reviewer; and is

The group of pathologists represents a subsequent group of reviewers.

53. The system of claim 49, wherein the instructions further program the processing unit to:

in response to a user requesting an interpretation of a particular WSI's transmission via a UI element:

generating an explanatory description using the standard pathology vocabulary and the stored explanatory information; and is

Displaying the explanatory description.

54. The system of claim 53, wherein the instructions further program the processing unit to:

selecting, from the particular WSI, an ROI for which the specified label indicates a severe condition or for which the confidence score is equal to or below a specified threshold confidence score;

highlighting in the ROI the feature of a particular feature type using a color assigned to the particular feature type, the particular feature type being indicative, at least in part, of the tissue condition indicated by the label assigned to the ROI; and is

Displaying the highlighted ROI along with the explanatory description.

55. The system of claim 49, wherein, for a first ROI in a first WSI, the explanatory information includes one or more of:

a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI;

a count of features of a particular feature type detected in the first ROI;

a measured density of features of the particular feature type in the first ROI; or

A strength of the particular feature type in indicating a corresponding tissue condition.

56. The system of claim 49, wherein the confidence score for a first ROI in a first WSI is based on one or more of:

a total number of feature types detected in the first ROI and corresponding to the tissue condition indicated by the label assigned to the first ROI;

for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of the first feature type detected in the ROI; or

Another total number of feature types detected in the first ROI but corresponding to a tissue condition different from the condition associated with the label assigned to the first ROI.

57. A system for truth labeling images used to train a classifier, the system comprising:

a first processor; and

a first memory in electrical communication with the first processor and comprising instructions that, when executed by a processing unit comprising the first or second processor and in electrical communication with a storage module comprising the first or second memory, program the processing unit to:

obtaining a full-section image (WSI) of the tissue;

identifying one or more regions of interest (ROIs) in the WSI, wherein identifying an ROI comprises: (i) marking at least two types of superpixels in the WSI, one type corresponding to hematoxylin stained tissue and the other type corresponding to eosin stained tissue; and (ii) marking segments of pixels of a first type to define an enclosed region as the ROI;

displaying one or more ROIs in sequence; and

for each ROI:

displaying one or more UI elements, a first UI element providing or confirming a respective truth label to be assigned to the ROI; and

in response to a user interaction using the first UI element, the respective truth label is specified to the ROI and the ROI is stored in a training corpus.

58. The system of claim 57, wherein,

the first UI element indicates agreement of the user with the provided suggestion, an

The instructions also program the processing unit to: for each ROI in at least a subset of the one or more ROIs:

identifying features of a plurality of feature types, wherein at least one feature type is indicative, at least in part, of a pathological condition of the tissue within the ROI;

operating as a classifier trained to classify an image into one of a plurality of categories of tissue conditions using features of a plurality of feature types to:

(i) classifying the ROI into a category within a plurality of categories;

(ii) assigning a suggestion tag to the ROI, the suggestion tag indicating a tissue condition associated with the category; and is

(iii) Storing the specified explanatory information about the suggested tag, the explanatory information including information about the identified features; and is

Displaying the suggestion tag as the provided suggestion.

59. The system of claim 58, wherein the instructions further program the processing unit to: in response to a user requesting interpretation of the suggested tags for a particular ROI via a UI element:

generating an explanatory description using the standard pathology vocabulary and the stored explanatory information; and is

Displaying the explanatory description.

60. The system of claim 59, wherein the instructions further program the processing unit to:

highlighting in the particular ROI the feature of a particular feature type using a color assigned to the particular feature type, the particular feature type being indicative, at least in part, of the tissue condition indicated by the label assigned to the ROI; and is

Displaying the highlighted ROI along with the explanatory description.

61. The system of claim 59, wherein for the particular ROI, the explanatory information includes one or more of:

a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI;

a count of features of a particular feature type detected in the particular ROI;

a measured density of features of the particular feature type in the particular ROI; or

A strength of the particular feature type in indicating a corresponding tissue condition.

62. The system of claim 59, wherein the confidence score for the particular ROI is based on one or more of:

a total number of feature types detected in the particular ROI and corresponding to the tissue condition indicated by the suggestion tag assigned to the particular ROI;

for a first feature type: (i) a strength of the first feature type in indicating a corresponding tissue condition; or (ii) a count of features of the first feature type detected in the particular ROI; or

Another total number of feature types detected in the particular ROI but corresponding to tissue conditions other than the condition associated with the suggestion tag assigned to the particular ROI.

Images